Upload 8 files

util/box_annotator.py

@@ -0,0 +1,262 @@
+from typing import List, Optional, Union, Tuple
+
+import cv2
+import numpy as np
+
+from supervision.detection.core import Detections
+from supervision.draw.color import Color, ColorPalette
+
+
+class BoxAnnotator:
+    """
+    A class for drawing bounding boxes on an image using detections provided.
+
+    Attributes:
+        color (Union[Color, ColorPalette]): The color to draw the bounding box,
+            can be a single color or a color palette
+        thickness (int): The thickness of the bounding box lines, default is 2
+        text_color (Color): The color of the text on the bounding box, default is white
+        text_scale (float): The scale of the text on the bounding box, default is 0.5
+        text_thickness (int): The thickness of the text on the bounding box,
+            default is 1
+        text_padding (int): The padding around the text on the bounding box,
+            default is 5
+
+    """
+
+    def __init__(
+        self,
+        color: Union[Color, ColorPalette] = ColorPalette.DEFAULT,
+        thickness: int = 3, # 1 for seeclick 2 for mind2web and 3 for demo
+        text_color: Color = Color.BLACK,
+        text_scale: float = 0.5, # 0.8 for mobile/web, 0.3 for desktop # 0.4 for mind2web
+        text_thickness: int = 2, #1, # 2 for demo
+        text_padding: int = 10,
+        avoid_overlap: bool = True,
+    ):
+        self.color: Union[Color, ColorPalette] = color
+        self.thickness: int = thickness
+        self.text_color: Color = text_color
+        self.text_scale: float = text_scale
+        self.text_thickness: int = text_thickness
+        self.text_padding: int = text_padding
+        self.avoid_overlap: bool = avoid_overlap
+
+    def annotate(
+        self,
+        scene: np.ndarray,
+        detections: Detections,
+        labels: Optional[List[str]] = None,
+        skip_label: bool = False,
+        image_size: Optional[Tuple[int, int]] = None,
+    ) -> np.ndarray:
+        """
+        Draws bounding boxes on the frame using the detections provided.
+
+        Args:
+            scene (np.ndarray): The image on which the bounding boxes will be drawn
+            detections (Detections): The detections for which the
+                bounding boxes will be drawn
+            labels (Optional[List[str]]): An optional list of labels
+                corresponding to each detection. If `labels` are not provided,
+                corresponding `class_id` will be used as label.
+            skip_label (bool): Is set to `True`, skips bounding box label annotation.
+        Returns:
+            np.ndarray: The image with the bounding boxes drawn on it
+
+        Example:
+            ```python
+            import supervision as sv
+
+            classes = ['person', ...]
+            image = ...
+            detections = sv.Detections(...)
+
+            box_annotator = sv.BoxAnnotator()
+            labels = [
+                f"{classes[class_id]} {confidence:0.2f}"
+                for _, _, confidence, class_id, _ in detections
+            ]
+            annotated_frame = box_annotator.annotate(
+                scene=image.copy(),
+                detections=detections,
+                labels=labels
+            )
+            ```
+        """
+        font = cv2.FONT_HERSHEY_SIMPLEX
+        for i in range(len(detections)):
+            x1, y1, x2, y2 = detections.xyxy[i].astype(int)
+            class_id = (
+                detections.class_id[i] if detections.class_id is not None else None
+            )
+            idx = class_id if class_id is not None else i
+            color = (
+                self.color.by_idx(idx)
+                if isinstance(self.color, ColorPalette)
+                else self.color
+            )
+            cv2.rectangle(
+                img=scene,
+                pt1=(x1, y1),
+                pt2=(x2, y2),
+                color=color.as_bgr(),
+                thickness=self.thickness,
+            )
+            if skip_label:
+                continue
+
+            text = (
+                f"{class_id}"
+                if (labels is None or len(detections) != len(labels))
+                else labels[i]
+            )
+
+            text_width, text_height = cv2.getTextSize(
+                text=text,
+                fontFace=font,
+                fontScale=self.text_scale,
+                thickness=self.text_thickness,
+            )[0]
+
+            if not self.avoid_overlap:
+                text_x = x1 + self.text_padding
+                text_y = y1 - self.text_padding
+
+                text_background_x1 = x1
+                text_background_y1 = y1 - 2 * self.text_padding - text_height
+
+                text_background_x2 = x1 + 2 * self.text_padding + text_width
+                text_background_y2 = y1
+                # text_x = x1 - self.text_padding - text_width
+                # text_y = y1 + self.text_padding + text_height
+                # text_background_x1 = x1 - 2 * self.text_padding - text_width
+                # text_background_y1 = y1
+                # text_background_x2 = x1
+                # text_background_y2 = y1 + 2 * self.text_padding + text_height
+            else:
+                text_x, text_y, text_background_x1, text_background_y1, text_background_x2, text_background_y2 = get_optimal_label_pos(self.text_padding, text_width, text_height, x1, y1, x2, y2, detections, image_size)
+
+            cv2.rectangle(
+                img=scene,
+                pt1=(text_background_x1, text_background_y1),
+                pt2=(text_background_x2, text_background_y2),
+                color=color.as_bgr(),
+                thickness=cv2.FILLED,
+            )
+            # import pdb; pdb.set_trace()
+            box_color = color.as_rgb()
+            luminance = 0.299 * box_color[0] + 0.587 * box_color[1] + 0.114 * box_color[2]
+            text_color = (0,0,0) if luminance > 160 else (255,255,255)
+            cv2.putText(
+                img=scene,
+                text=text,
+                org=(text_x, text_y),
+                fontFace=font,
+                fontScale=self.text_scale,
+                # color=self.text_color.as_rgb(),
+                color=text_color,
+                thickness=self.text_thickness,
+                lineType=cv2.LINE_AA,
+            )
+        return scene
+    
+
+def box_area(box):
+        return (box[2] - box[0]) * (box[3] - box[1])
+
+def intersection_area(box1, box2):
+    x1 = max(box1[0], box2[0])
+    y1 = max(box1[1], box2[1])
+    x2 = min(box1[2], box2[2])
+    y2 = min(box1[3], box2[3])
+    return max(0, x2 - x1) * max(0, y2 - y1)
+
+def IoU(box1, box2, return_max=True):
+    intersection = intersection_area(box1, box2)
+    union = box_area(box1) + box_area(box2) - intersection
+    if box_area(box1) > 0 and box_area(box2) > 0:
+        ratio1 = intersection / box_area(box1)
+        ratio2 = intersection / box_area(box2)
+    else:
+        ratio1, ratio2 = 0, 0
+    if return_max:
+        return max(intersection / union, ratio1, ratio2)
+    else:
+        return intersection / union
+
+
+def get_optimal_label_pos(text_padding, text_width, text_height, x1, y1, x2, y2, detections, image_size):
+    """ check overlap of text and background detection box, and get_optimal_label_pos, 
+        pos: str, position of the text, must be one of 'top left', 'top right', 'outer left', 'outer right' TODO: if all are overlapping, return the last one, i.e. outer right
+        Threshold: default to 0.3
+    """
+
+    def get_is_overlap(detections, text_background_x1, text_background_y1, text_background_x2, text_background_y2, image_size):
+        is_overlap = False
+        for i in range(len(detections)):
+            detection = detections.xyxy[i].astype(int)
+            if IoU([text_background_x1, text_background_y1, text_background_x2, text_background_y2], detection) > 0.3:
+                is_overlap = True
+                break
+        # check if the text is out of the image
+        if text_background_x1 < 0 or text_background_x2 > image_size[0] or text_background_y1 < 0 or text_background_y2 > image_size[1]:
+            is_overlap = True
+        return is_overlap
+    
+    # if pos == 'top left':
+    text_x = x1 + text_padding
+    text_y = y1 - text_padding
+
+    text_background_x1 = x1
+    text_background_y1 = y1 - 2 * text_padding - text_height
+
+    text_background_x2 = x1 + 2 * text_padding + text_width
+    text_background_y2 = y1
+    is_overlap = get_is_overlap(detections, text_background_x1, text_background_y1, text_background_x2, text_background_y2, image_size)
+    if not is_overlap:
+        return text_x, text_y, text_background_x1, text_background_y1, text_background_x2, text_background_y2
+    
+    # elif pos == 'outer left':
+    text_x = x1 - text_padding - text_width
+    text_y = y1 + text_padding + text_height
+
+    text_background_x1 = x1 - 2 * text_padding - text_width
+    text_background_y1 = y1
+
+    text_background_x2 = x1
+    text_background_y2 = y1 + 2 * text_padding + text_height
+    is_overlap = get_is_overlap(detections, text_background_x1, text_background_y1, text_background_x2, text_background_y2, image_size)
+    if not is_overlap:
+        return text_x, text_y, text_background_x1, text_background_y1, text_background_x2, text_background_y2
+    
+
+    # elif pos == 'outer right':
+    text_x = x2 + text_padding
+    text_y = y1 + text_padding + text_height
+
+    text_background_x1 = x2
+    text_background_y1 = y1
+
+    text_background_x2 = x2 + 2 * text_padding + text_width
+    text_background_y2 = y1 + 2 * text_padding + text_height
+
+    is_overlap = get_is_overlap(detections, text_background_x1, text_background_y1, text_background_x2, text_background_y2, image_size)
+    if not is_overlap:
+        return text_x, text_y, text_background_x1, text_background_y1, text_background_x2, text_background_y2
+
+    # elif pos == 'top right':
+    text_x = x2 - text_padding - text_width
+    text_y = y1 - text_padding
+
+    text_background_x1 = x2 - 2 * text_padding - text_width
+    text_background_y1 = y1 - 2 * text_padding - text_height
+
+    text_background_x2 = x2
+    text_background_y2 = y1
+
+    is_overlap = get_is_overlap(detections, text_background_x1, text_background_y1, text_background_x2, text_background_y2, image_size)
+    if not is_overlap:
+        return text_x, text_y, text_background_x1, text_background_y1, text_background_x2, text_background_y2
+
+    return text_x, text_y, text_background_x1, text_background_y1, text_background_x2, text_background_y2

util/omniparser.py

@@ -0,0 +1,32 @@
+from util.utils import get_som_labeled_img, get_caption_model_processor, get_yolo_model, check_ocr_box
+import torch
+from PIL import Image
+import io
+import base64
+from typing import Dict
+class Omniparser(object):
+    def __init__(self, config: Dict):
+        self.config = config
+        device = 'cuda' if torch.cuda.is_available() else 'cpu'
+
+        self.som_model = get_yolo_model(model_path=config['som_model_path'])
+        self.caption_model_processor = get_caption_model_processor(model_name=config['caption_model_name'], model_name_or_path=config['caption_model_path'], device=device)
+        print('Omniparser initialized!!!')
+
+    def parse(self, image_base64: str):
+        image_bytes = base64.b64decode(image_base64)
+        image = Image.open(io.BytesIO(image_bytes))
+        print('image size:', image.size)
+        
+        box_overlay_ratio = max(image.size) / 3200
+        draw_bbox_config = {
+            'text_scale': 0.8 * box_overlay_ratio,
+            'text_thickness': max(int(2 * box_overlay_ratio), 1),
+            'text_padding': max(int(3 * box_overlay_ratio), 1),
+            'thickness': max(int(3 * box_overlay_ratio), 1),
+        }
+
+        (text, ocr_bbox), _ = check_ocr_box(image, display_img=False, output_bb_format='xyxy', easyocr_args={'text_threshold': 0.8}, use_paddleocr=False)
+        dino_labled_img, label_coordinates, parsed_content_list = get_som_labeled_img(image, self.som_model, BOX_TRESHOLD = self.config['BOX_TRESHOLD'], output_coord_in_ratio=True, ocr_bbox=ocr_bbox,draw_bbox_config=draw_bbox_config, caption_model_processor=self.caption_model_processor, ocr_text=text,use_local_semantics=True, iou_threshold=0.7, scale_img=False, batch_size=128)
+
+        return dino_labled_img, parsed_content_list

util/utils - 副本.py

@@ -0,0 +1,540 @@
+# from ultralytics import YOLO
+import os
+import io
+import base64
+import time
+from PIL import Image, ImageDraw, ImageFont
+import json
+import requests
+# utility function
+import os
+from openai import AzureOpenAI
+
+import json
+import sys
+import os
+import cv2
+import numpy as np
+# %matplotlib inline
+from matplotlib import pyplot as plt
+import easyocr
+from paddleocr import PaddleOCR
+reader = easyocr.Reader(['en'])
+paddle_ocr = PaddleOCR(
+    lang='en',  # other lang also available
+    use_angle_cls=False,
+    use_gpu=False,  # using cuda will conflict with pytorch in the same process
+    show_log=False,
+    max_batch_size=1024,
+    use_dilation=True,  # improves accuracy
+    det_db_score_mode='slow',  # improves accuracy
+    rec_batch_num=1024)
+import time
+import base64
+
+import os
+import ast
+import torch
+from typing import Tuple, List, Union
+from torchvision.ops import box_convert
+import re
+from torchvision.transforms import ToPILImage
+import supervision as sv
+import torchvision.transforms as T
+from util.box_annotator import BoxAnnotator 
+
+
+def get_caption_model_processor(model_name, model_name_or_path="Salesforce/blip2-opt-2.7b", device=None):
+    if not device:
+        device = "cuda" if torch.cuda.is_available() else "cpu"
+    if model_name == "blip2":
+        from transformers import Blip2Processor, Blip2ForConditionalGeneration
+        processor = Blip2Processor.from_pretrained("Salesforce/blip2-opt-2.7b")
+        if device == 'cpu':
+            model = Blip2ForConditionalGeneration.from_pretrained(
+            model_name_or_path, device_map=None, torch_dtype=torch.float32
+        ) 
+        else:
+            model = Blip2ForConditionalGeneration.from_pretrained(
+            model_name_or_path, device_map=None, torch_dtype=torch.float16
+        ).to(device)
+    elif model_name == "florence2":
+        from transformers import AutoProcessor, AutoModelForCausalLM 
+        processor = AutoProcessor.from_pretrained("microsoft/Florence-2-base", trust_remote_code=True)
+        if device == 'cpu':
+            model = AutoModelForCausalLM.from_pretrained(model_name_or_path, torch_dtype=torch.float32, trust_remote_code=True)
+        else:
+            model = AutoModelForCausalLM.from_pretrained(model_name_or_path, torch_dtype=torch.float16, trust_remote_code=True).to(device)
+    return {'model': model.to(device), 'processor': processor}
+
+
+def get_yolo_model(model_path):
+    from ultralytics import YOLO
+    # Load the model.
+    model = YOLO(model_path)
+    return model
+
+
+@torch.inference_mode()
+def get_parsed_content_icon(filtered_boxes, starting_idx, image_source, caption_model_processor, prompt=None, batch_size=128):
+    # Number of samples per batch, --> 128 roughly takes 4 GB of GPU memory for florence v2 model
+    to_pil = ToPILImage()
+    if starting_idx:
+        non_ocr_boxes = filtered_boxes[starting_idx:]
+    else:
+        non_ocr_boxes = filtered_boxes
+    croped_pil_image = []
+    for i, coord in enumerate(non_ocr_boxes):
+        try:
+            xmin, xmax = int(coord[0]*image_source.shape[1]), int(coord[2]*image_source.shape[1])
+            ymin, ymax = int(coord[1]*image_source.shape[0]), int(coord[3]*image_source.shape[0])
+            cropped_image = image_source[ymin:ymax, xmin:xmax, :]
+            cropped_image = cv2.resize(cropped_image, (64, 64))
+            croped_pil_image.append(to_pil(cropped_image))
+        except:
+            continue
+
+    model, processor = caption_model_processor['model'], caption_model_processor['processor']
+    if not prompt:
+        if 'florence' in model.config.name_or_path:
+            prompt = "<CAPTION>"
+        else:
+            prompt = "The image shows"
+    
+    generated_texts = []
+    device = model.device
+    for i in range(0, len(croped_pil_image), batch_size):
+        start = time.time()
+        batch = croped_pil_image[i:i+batch_size]
+        t1 = time.time()
+        if model.device.type == 'cuda':
+            inputs = processor(images=batch, text=[prompt]*len(batch), return_tensors="pt", do_resize=False).to(device=device, dtype=torch.float16)
+        else:
+            inputs = processor(images=batch, text=[prompt]*len(batch), return_tensors="pt").to(device=device)
+        if 'florence' in model.config.name_or_path:
+            generated_ids = model.generate(input_ids=inputs["input_ids"],pixel_values=inputs["pixel_values"],max_new_tokens=20,num_beams=1, do_sample=False)
+        else:
+            generated_ids = model.generate(**inputs, max_length=100, num_beams=5, no_repeat_ngram_size=2, early_stopping=True, num_return_sequences=1) # temperature=0.01, do_sample=True,
+        generated_text = processor.batch_decode(generated_ids, skip_special_tokens=True)
+        generated_text = [gen.strip() for gen in generated_text]
+        generated_texts.extend(generated_text)
+    
+    return generated_texts
+
+
+
+def get_parsed_content_icon_phi3v(filtered_boxes, ocr_bbox, image_source, caption_model_processor):
+    to_pil = ToPILImage()
+    if ocr_bbox:
+        non_ocr_boxes = filtered_boxes[len(ocr_bbox):]
+    else:
+        non_ocr_boxes = filtered_boxes
+    croped_pil_image = []
+    for i, coord in enumerate(non_ocr_boxes):
+        xmin, xmax = int(coord[0]*image_source.shape[1]), int(coord[2]*image_source.shape[1])
+        ymin, ymax = int(coord[1]*image_source.shape[0]), int(coord[3]*image_source.shape[0])
+        cropped_image = image_source[ymin:ymax, xmin:xmax, :]
+        croped_pil_image.append(to_pil(cropped_image))
+
+    model, processor = caption_model_processor['model'], caption_model_processor['processor']
+    device = model.device
+    messages = [{"role": "user", "content": "<|image_1|>\ndescribe the icon in one sentence"}] 
+    prompt = processor.tokenizer.apply_chat_template(messages, tokenize=False, add_generation_prompt=True)
+
+    batch_size = 5  # Number of samples per batch
+    generated_texts = []
+
+    for i in range(0, len(croped_pil_image), batch_size):
+        images = croped_pil_image[i:i+batch_size]
+        image_inputs = [processor.image_processor(x, return_tensors="pt") for x in images]
+        inputs ={'input_ids': [], 'attention_mask': [], 'pixel_values': [], 'image_sizes': []}
+        texts = [prompt] * len(images)
+        for i, txt in enumerate(texts):
+            input = processor._convert_images_texts_to_inputs(image_inputs[i], txt, return_tensors="pt")
+            inputs['input_ids'].append(input['input_ids'])
+            inputs['attention_mask'].append(input['attention_mask'])
+            inputs['pixel_values'].append(input['pixel_values'])
+            inputs['image_sizes'].append(input['image_sizes'])
+        max_len = max([x.shape[1] for x in inputs['input_ids']])
+        for i, v in enumerate(inputs['input_ids']):
+            inputs['input_ids'][i] = torch.cat([processor.tokenizer.pad_token_id * torch.ones(1, max_len - v.shape[1], dtype=torch.long), v], dim=1)
+            inputs['attention_mask'][i] = torch.cat([torch.zeros(1, max_len - v.shape[1], dtype=torch.long), inputs['attention_mask'][i]], dim=1)
+        inputs_cat = {k: torch.concatenate(v).to(device) for k, v in inputs.items()}
+
+        generation_args = { 
+            "max_new_tokens": 25, 
+            "temperature": 0.01, 
+            "do_sample": False, 
+        } 
+        generate_ids = model.generate(**inputs_cat, eos_token_id=processor.tokenizer.eos_token_id, **generation_args) 
+        # # remove input tokens 
+        generate_ids = generate_ids[:, inputs_cat['input_ids'].shape[1]:]
+        response = processor.batch_decode(generate_ids, skip_special_tokens=True, clean_up_tokenization_spaces=False)
+        response = [res.strip('\n').strip() for res in response]
+        generated_texts.extend(response)
+
+    return generated_texts
+
+def remove_overlap(boxes, iou_threshold, ocr_bbox=None):
+    assert ocr_bbox is None or isinstance(ocr_bbox, List)
+
+    def box_area(box):
+        return (box[2] - box[0]) * (box[3] - box[1])
+
+    def intersection_area(box1, box2):
+        x1 = max(box1[0], box2[0])
+        y1 = max(box1[1], box2[1])
+        x2 = min(box1[2], box2[2])
+        y2 = min(box1[3], box2[3])
+        return max(0, x2 - x1) * max(0, y2 - y1)
+
+    def IoU(box1, box2):
+        intersection = intersection_area(box1, box2)
+        union = box_area(box1) + box_area(box2) - intersection + 1e-6
+        if box_area(box1) > 0 and box_area(box2) > 0:
+            ratio1 = intersection / box_area(box1)
+            ratio2 = intersection / box_area(box2)
+        else:
+            ratio1, ratio2 = 0, 0
+        return max(intersection / union, ratio1, ratio2)
+
+    def is_inside(box1, box2):
+        # return box1[0] >= box2[0] and box1[1] >= box2[1] and box1[2] <= box2[2] and box1[3] <= box2[3]
+        intersection = intersection_area(box1, box2)
+        ratio1 = intersection / box_area(box1)
+        return ratio1 > 0.95
+
+    boxes = boxes.tolist()
+    filtered_boxes = []
+    if ocr_bbox:
+        filtered_boxes.extend(ocr_bbox)
+    # print('ocr_bbox!!!', ocr_bbox)
+    for i, box1 in enumerate(boxes):
+        # if not any(IoU(box1, box2) > iou_threshold and box_area(box1) > box_area(box2) for j, box2 in enumerate(boxes) if i != j):
+        is_valid_box = True
+        for j, box2 in enumerate(boxes):
+            # keep the smaller box
+            if i != j and IoU(box1, box2) > iou_threshold and box_area(box1) > box_area(box2):
+                is_valid_box = False
+                break
+        if is_valid_box:
+            # add the following 2 lines to include ocr bbox
+            if ocr_bbox:
+                # only add the box if it does not overlap with any ocr bbox
+                if not any(IoU(box1, box3) > iou_threshold and not is_inside(box1, box3) for k, box3 in enumerate(ocr_bbox)):
+                    filtered_boxes.append(box1)
+            else:
+                filtered_boxes.append(box1)
+    return torch.tensor(filtered_boxes)
+
+
+def remove_overlap_new(boxes, iou_threshold, ocr_bbox=None):
+    '''
+    ocr_bbox format: [{'type': 'text', 'bbox':[x,y], 'interactivity':False, 'content':str }, ...]
+    boxes format: [{'type': 'icon', 'bbox':[x,y], 'interactivity':True, 'content':None }, ...]
+
+    '''
+    assert ocr_bbox is None or isinstance(ocr_bbox, List)
+
+    def box_area(box):
+        return (box[2] - box[0]) * (box[3] - box[1])
+
+    def intersection_area(box1, box2):
+        x1 = max(box1[0], box2[0])
+        y1 = max(box1[1], box2[1])
+        x2 = min(box1[2], box2[2])
+        y2 = min(box1[3], box2[3])
+        return max(0, x2 - x1) * max(0, y2 - y1)
+
+    def IoU(box1, box2):
+        intersection = intersection_area(box1, box2)
+        union = box_area(box1) + box_area(box2) - intersection + 1e-6
+        if box_area(box1) > 0 and box_area(box2) > 0:
+            ratio1 = intersection / box_area(box1)
+            ratio2 = intersection / box_area(box2)
+        else:
+            ratio1, ratio2 = 0, 0
+        return max(intersection / union, ratio1, ratio2)
+
+    def is_inside(box1, box2):
+        # return box1[0] >= box2[0] and box1[1] >= box2[1] and box1[2] <= box2[2] and box1[3] <= box2[3]
+        intersection = intersection_area(box1, box2)
+        ratio1 = intersection / box_area(box1)
+        return ratio1 > 0.80
+
+    # boxes = boxes.tolist()
+    filtered_boxes = []
+    if ocr_bbox:
+        filtered_boxes.extend(ocr_bbox)
+    # print('ocr_bbox!!!', ocr_bbox)
+    for i, box1_elem in enumerate(boxes):
+        box1 = box1_elem['bbox']
+        is_valid_box = True
+        for j, box2_elem in enumerate(boxes):
+            # keep the smaller box
+            box2 = box2_elem['bbox']
+            if i != j and IoU(box1, box2) > iou_threshold and box_area(box1) > box_area(box2):
+                is_valid_box = False
+                break
+        if is_valid_box:
+            if ocr_bbox:
+                # keep yolo boxes + prioritize ocr label
+                box_added = False
+                ocr_labels = ''
+                for box3_elem in ocr_bbox:
+                    if not box_added:
+                        box3 = box3_elem['bbox']
+                        if is_inside(box3, box1): # ocr inside icon
+                            # box_added = True
+                            # delete the box3_elem from ocr_bbox
+                            try:
+                                # gather all ocr labels
+                                ocr_labels += box3_elem['content'] + ' '
+                                filtered_boxes.remove(box3_elem)
+                            except:
+                                continue
+                            # break
+                        elif is_inside(box1, box3): # icon inside ocr, don't added this icon box, no need to check other ocr bbox bc no overlap between ocr bbox, icon can only be in one ocr box
+                            box_added = True
+                            break
+                        else:
+                            continue
+                if not box_added:
+                    if ocr_labels:
+                        filtered_boxes.append({'type': 'icon', 'bbox': box1_elem['bbox'], 'interactivity': True, 'content': ocr_labels, 'source':'box_yolo_content_ocr'})
+                    else:
+                        filtered_boxes.append({'type': 'icon', 'bbox': box1_elem['bbox'], 'interactivity': True, 'content': None, 'source':'box_yolo_content_yolo'})
+            else:
+                filtered_boxes.append(box1)
+    return filtered_boxes # torch.tensor(filtered_boxes)
+
+
+def load_image(image_path: str) -> Tuple[np.array, torch.Tensor]:
+    transform = T.Compose(
+        [
+            T.RandomResize([800], max_size=1333),
+            T.ToTensor(),
+            T.Normalize([0.485, 0.456, 0.406], [0.229, 0.224, 0.225]),
+        ]
+    )
+    image_source = Image.open(image_path).convert("RGB")
+    image = np.asarray(image_source)
+    image_transformed, _ = transform(image_source, None)
+    return image, image_transformed
+
+
+def annotate(image_source: np.ndarray, boxes: torch.Tensor, logits: torch.Tensor, phrases: List[str], text_scale: float, 
+             text_padding=5, text_thickness=2, thickness=3) -> np.ndarray:
+    """    
+    This function annotates an image with bounding boxes and labels.
+
+    Parameters:
+    image_source (np.ndarray): The source image to be annotated.
+    boxes (torch.Tensor): A tensor containing bounding box coordinates. in cxcywh format, pixel scale
+    logits (torch.Tensor): A tensor containing confidence scores for each bounding box.
+    phrases (List[str]): A list of labels for each bounding box.
+    text_scale (float): The scale of the text to be displayed. 0.8 for mobile/web, 0.3 for desktop # 0.4 for mind2web
+
+    Returns:
+    np.ndarray: The annotated image.
+    """
+    h, w, _ = image_source.shape
+    boxes = boxes * torch.Tensor([w, h, w, h])
+    xyxy = box_convert(boxes=boxes, in_fmt="cxcywh", out_fmt="xyxy").numpy()
+    xywh = box_convert(boxes=boxes, in_fmt="cxcywh", out_fmt="xywh").numpy()
+    detections = sv.Detections(xyxy=xyxy)
+
+    labels = [f"{phrase}" for phrase in range(boxes.shape[0])]
+
+    box_annotator = BoxAnnotator(text_scale=text_scale, text_padding=text_padding,text_thickness=text_thickness,thickness=thickness) # 0.8 for mobile/web, 0.3 for desktop # 0.4 for mind2web
+    annotated_frame = image_source.copy()
+    annotated_frame = box_annotator.annotate(scene=annotated_frame, detections=detections, labels=labels, image_size=(w,h))
+
+    label_coordinates = {f"{phrase}": v for phrase, v in zip(phrases, xywh)}
+    return annotated_frame, label_coordinates
+
+
+def predict(model, image, caption, box_threshold, text_threshold):
+    """ Use huggingface model to replace the original model
+    """
+    model, processor = model['model'], model['processor']
+    device = model.device
+
+    inputs = processor(images=image, text=caption, return_tensors="pt").to(device)
+    with torch.no_grad():
+        outputs = model(**inputs)
+
+    results = processor.post_process_grounded_object_detection(
+        outputs,
+        inputs.input_ids,
+        box_threshold=box_threshold, # 0.4,
+        text_threshold=text_threshold, # 0.3,
+        target_sizes=[image.size[::-1]]
+    )[0]
+    boxes, logits, phrases = results["boxes"], results["scores"], results["labels"]
+    return boxes, logits, phrases
+
+
+def predict_yolo(model, image, box_threshold, imgsz, scale_img, iou_threshold=0.7):
+    """ Use huggingface model to replace the original model
+    """
+    # model = model['model']
+    if scale_img:
+        result = model.predict(
+        source=image,
+        conf=box_threshold,
+        imgsz=imgsz,
+        iou=iou_threshold, # default 0.7
+        )
+    else:
+        result = model.predict(
+        source=image,
+        conf=box_threshold,
+        iou=iou_threshold, # default 0.7
+        )
+    boxes = result[0].boxes.xyxy#.tolist() # in pixel space
+    conf = result[0].boxes.conf
+    phrases = [str(i) for i in range(len(boxes))]
+
+    return boxes, conf, phrases
+
+def int_box_area(box, w, h):
+    x1, y1, x2, y2 = box
+    int_box = [int(x1*w), int(y1*h), int(x2*w), int(y2*h)]
+    area = (int_box[2] - int_box[0]) * (int_box[3] - int_box[1])
+    return area
+
+def get_som_labeled_img(image_source: Union[str, Image.Image], model=None, BOX_TRESHOLD=0.01, output_coord_in_ratio=False, ocr_bbox=None, text_scale=0.4, text_padding=5, draw_bbox_config=None, caption_model_processor=None, ocr_text=[], use_local_semantics=True, iou_threshold=0.9,prompt=None, scale_img=False, imgsz=None, batch_size=128):
+    """Process either an image path or Image object
+    
+    Args:
+        image_source: Either a file path (str) or PIL Image object
+        ...
+    """
+    if isinstance(image_source, str):
+        image_source = Image.open(image_source)
+    image_source = image_source.convert("RGB") # for CLIP
+    w, h = image_source.size
+    if not imgsz:
+        imgsz = (h, w)
+    # print('image size:', w, h)
+    xyxy, logits, phrases = predict_yolo(model=model, image=image_source, box_threshold=BOX_TRESHOLD, imgsz=imgsz, scale_img=scale_img, iou_threshold=0.1)
+    xyxy = xyxy / torch.Tensor([w, h, w, h]).to(xyxy.device)
+    image_source = np.asarray(image_source)
+    phrases = [str(i) for i in range(len(phrases))]
+
+    # annotate the image with labels
+    if ocr_bbox:
+        ocr_bbox = torch.tensor(ocr_bbox) / torch.Tensor([w, h, w, h])
+        ocr_bbox=ocr_bbox.tolist()
+    else:
+        print('no ocr bbox!!!')
+        ocr_bbox = None
+
+    ocr_bbox_elem = [{'type': 'text', 'bbox':box, 'interactivity':False, 'content':txt, 'source': 'box_ocr_content_ocr'} for box, txt in zip(ocr_bbox, ocr_text) if int_box_area(box, w, h) > 0] 
+    xyxy_elem = [{'type': 'icon', 'bbox':box, 'interactivity':True, 'content':None} for box in xyxy.tolist() if int_box_area(box, w, h) > 0]
+    filtered_boxes = remove_overlap_new(boxes=xyxy_elem, iou_threshold=iou_threshold, ocr_bbox=ocr_bbox_elem)
+    
+    # sort the filtered_boxes so that the one with 'content': None is at the end, and get the index of the first 'content': None
+    filtered_boxes_elem = sorted(filtered_boxes, key=lambda x: x['content'] is None)
+    # get the index of the first 'content': None
+    starting_idx = next((i for i, box in enumerate(filtered_boxes_elem) if box['content'] is None), -1)
+    filtered_boxes = torch.tensor([box['bbox'] for box in filtered_boxes_elem])
+    print('len(filtered_boxes):', len(filtered_boxes), starting_idx)
+
+    # get parsed icon local semantics
+    time1 = time.time()
+    if use_local_semantics:
+        caption_model = caption_model_processor['model']
+        if 'phi3_v' in caption_model.config.model_type: 
+            parsed_content_icon = get_parsed_content_icon_phi3v(filtered_boxes, ocr_bbox, image_source, caption_model_processor)
+        else:
+            parsed_content_icon = get_parsed_content_icon(filtered_boxes, starting_idx, image_source, caption_model_processor, prompt=prompt,batch_size=batch_size)
+        ocr_text = [f"Text Box ID {i}: {txt}" for i, txt in enumerate(ocr_text)]
+        icon_start = len(ocr_text)
+        parsed_content_icon_ls = []
+        # fill the filtered_boxes_elem None content with parsed_content_icon in order
+        for i, box in enumerate(filtered_boxes_elem):
+            if box['content'] is None:
+                box['content'] = parsed_content_icon.pop(0)
+        for i, txt in enumerate(parsed_content_icon):
+            parsed_content_icon_ls.append(f"Icon Box ID {str(i+icon_start)}: {txt}")
+        parsed_content_merged = ocr_text + parsed_content_icon_ls
+    else:
+        ocr_text = [f"Text Box ID {i}: {txt}" for i, txt in enumerate(ocr_text)]
+        parsed_content_merged = ocr_text
+    print('time to get parsed content:', time.time()-time1)
+
+    filtered_boxes = box_convert(boxes=filtered_boxes, in_fmt="xyxy", out_fmt="cxcywh")
+
+    phrases = [i for i in range(len(filtered_boxes))]
+    
+    # draw boxes
+    if draw_bbox_config:
+        annotated_frame, label_coordinates = annotate(image_source=image_source, boxes=filtered_boxes, logits=logits, phrases=phrases, **draw_bbox_config)
+    else:
+        annotated_frame, label_coordinates = annotate(image_source=image_source, boxes=filtered_boxes, logits=logits, phrases=phrases, text_scale=text_scale, text_padding=text_padding)
+    
+    pil_img = Image.fromarray(annotated_frame)
+    buffered = io.BytesIO()
+    pil_img.save(buffered, format="PNG")
+    encoded_image = base64.b64encode(buffered.getvalue()).decode('ascii')
+    if output_coord_in_ratio:
+        label_coordinates = {k: [v[0]/w, v[1]/h, v[2]/w, v[3]/h] for k, v in label_coordinates.items()}
+        assert w == annotated_frame.shape[1] and h == annotated_frame.shape[0]
+
+    return encoded_image, label_coordinates, filtered_boxes_elem
+
+
+def get_xywh(input):
+    x, y, w, h = input[0][0], input[0][1], input[2][0] - input[0][0], input[2][1] - input[0][1]
+    x, y, w, h = int(x), int(y), int(w), int(h)
+    return x, y, w, h
+
+def get_xyxy(input):
+    x, y, xp, yp = input[0][0], input[0][1], input[2][0], input[2][1]
+    x, y, xp, yp = int(x), int(y), int(xp), int(yp)
+    return x, y, xp, yp
+
+def get_xywh_yolo(input):
+    x, y, w, h = input[0], input[1], input[2] - input[0], input[3] - input[1]
+    x, y, w, h = int(x), int(y), int(w), int(h)
+    return x, y, w, h
+
+def check_ocr_box(image_source: Union[str, Image.Image], display_img = True, output_bb_format='xywh', goal_filtering=None, easyocr_args=None, use_paddleocr=False):
+    if isinstance(image_source, str):
+        image_source = Image.open(image_source)
+    if image_source.mode == 'RGBA':
+        # Convert RGBA to RGB to avoid alpha channel issues
+        image_source = image_source.convert('RGB')
+    image_np = np.array(image_source)
+    w, h = image_source.size
+    if use_paddleocr:
+        if easyocr_args is None:
+            text_threshold = 0.5
+        else:
+            text_threshold = easyocr_args['text_threshold']
+        result = paddle_ocr.ocr(image_np, cls=False)[0]
+        coord = [item[0] for item in result if item[1][1] > text_threshold]
+        text = [item[1][0] for item in result if item[1][1] > text_threshold]
+    else:  # EasyOCR
+        if easyocr_args is None:
+            easyocr_args = {}
+        result = reader.readtext(image_np, **easyocr_args)
+        coord = [item[0] for item in result]
+        text = [item[1] for item in result]
+    if display_img:
+        opencv_img = cv2.cvtColor(image_np, cv2.COLOR_RGB2BGR)
+        bb = []
+        for item in coord:
+            x, y, a, b = get_xywh(item)
+            bb.append((x, y, a, b))
+            cv2.rectangle(opencv_img, (x, y), (x+a, y+b), (0, 255, 0), 2)
+        #  matplotlib expects RGB
+        plt.imshow(cv2.cvtColor(opencv_img, cv2.COLOR_BGR2RGB))
+    else:
+        if output_bb_format == 'xywh':
+            bb = [get_xywh(item) for item in coord]
+        elif output_bb_format == 'xyxy':
+            bb = [get_xyxy(item) for item in coord]
+    return (text, bb), goal_filtering

util/utils.py

@@ -0,0 +1,641 @@
+# from ultralytics import YOLO
+import os
+import io
+import base64
+import time
+from PIL import Image, ImageDraw, ImageFont
+import json
+import requests
+import ollama
+# utility function
+import os
+from openai import AzureOpenAI
+
+import json
+import sys
+import os
+import cv2
+import numpy as np
+# %matplotlib inline
+from matplotlib import pyplot as plt
+import easyocr
+from paddleocr import PaddleOCR
+reader = easyocr.Reader(['en'])
+paddle_ocr = PaddleOCR(
+    lang='en',  # other lang also available
+    use_angle_cls=False,
+    use_gpu=False,  # using cuda will conflict with pytorch in the same process
+    show_log=False,
+    max_batch_size=1024,
+    use_dilation=True,  # improves accuracy
+    det_db_score_mode='slow',  # improves accuracy
+    rec_batch_num=1024)
+import time
+import base64
+
+import os
+import ast
+import torch
+from typing import Tuple, List, Union
+from torchvision.ops import box_convert
+import re
+from torchvision.transforms import ToPILImage
+import supervision as sv
+import torchvision.transforms as T
+from util.box_annotator import BoxAnnotator 
+
+
+def get_caption_model_processor(model_name, model_name_or_path="Salesforce/blip2-opt-2.7b", device=None):
+    if not device:
+        device = "cuda" if torch.cuda.is_available() else "cpu"
+    if model_name == "blip2":
+        from transformers import Blip2Processor, Blip2ForConditionalGeneration
+        processor = Blip2Processor.from_pretrained("Salesforce/blip2-opt-2.7b")
+        if device == 'cpu':
+            model = Blip2ForConditionalGeneration.from_pretrained(
+            model_name_or_path, device_map=None, torch_dtype=torch.float32
+        )
+        else:
+            model = Blip2ForConditionalGeneration.from_pretrained(
+            model_name_or_path, device_map=None, torch_dtype=torch.float16
+        ).to(device)
+    elif model_name == "florence":
+        from transformers import AutoProcessor, AutoModelForCausalLM
+        processor = AutoProcessor.from_pretrained("microsoft/Florence-base", trust_remote_code=True)
+        if device == 'cpu':
+            model = AutoModelForCausalLM.from_pretrained("microsoft/Florence-base", torch_dtype=torch.float32, trust_remote_code=True)
+        else:
+            model = AutoModelForCausalLM.from_pretrained("microsoft/Florence-base", torch_dtype=torch.float16, trust_remote_code=True).to(device)
+    elif model_name == "ollama":
+        return {"model": None, "processor": None}
+    elif model_name == "florence2":
+        from transformers import AutoProcessor, AutoModelForCausalLM
+        processor = AutoProcessor.from_pretrained("microsoft/Florence-2-base", trust_remote_code=True)
+        if device == 'cpu':
+            model = AutoModelForCausalLM.from_pretrained(model_name_or_path, torch_dtype=torch.float32, trust_remote_code=True)
+        else:
+            model = AutoModelForCausalLM.from_pretrained(model_name_or_path, torch_dtype=torch.float16, trust_remote_code=True).to(device)
+    return {'model': model.to(device), 'processor': processor}
+
+
+def get_yolo_model(model_path):
+    from ultralytics import YOLO
+    # Load the model.
+    model = YOLO(model_path)
+    return model
+
+
+@torch.inference_mode()
+def get_parsed_content_icon(filtered_boxes, starting_idx, image_source, caption_model_processor, prompt=None, batch_size=128):
+    parsed_content_icon = []
+    # Number of samples per batch, --> 128 roughly takes 4 GB of GPU memory for florence v2 model
+    to_pil = ToPILImage()
+    if starting_idx:
+        non_ocr_boxes = filtered_boxes[starting_idx:]
+    else:
+        non_ocr_boxes = filtered_boxes
+    croped_pil_image = []
+    for i, coord in enumerate(non_ocr_boxes):
+        try:
+            xmin, xmax = int(coord[0]*image_source.shape[1]), int(coord[2]*image_source.shape[1])
+            ymin, ymax = int(coord[1]*image_source.shape[0]), int(coord[3]*image_source.shape[0])
+            cropped_image = image_source[ymin:ymax, xmin:xmax, :]
+            cropped_image = cv2.resize(cropped_image, (64, 64))
+            croped_pil_image.append(to_pil(cropped_image))
+        except:
+            continue
+
+    model, processor = caption_model_processor['model'], caption_model_processor['processor']
+    if not prompt:
+        if 'florence' in model.config.name_or_path:
+            prompt = "<CAPTION>"
+        else:
+            prompt = "The image shows"
+    
+    generated_texts = []
+    device = model.device
+    for i in range(0, len(croped_pil_image), batch_size):
+        start = time.time()
+        batch = croped_pil_image[i:i+batch_size]
+        t1 = time.time()
+        if model.device.type == 'cuda':
+            inputs = processor(images=batch, text=[prompt]*len(batch), return_tensors="pt", do_resize=False).to(device=device, dtype=torch.float16)
+        else:
+            inputs = processor(images=batch, text=[prompt]*len(batch), return_tensors="pt").to(device=device)
+
+        # 显式生成 input_ids
+        text_inputs = processor.tokenizer([prompt]*len(batch), return_tensors="pt", padding=True)
+        inputs['input_ids'] = text_inputs['input_ids'].to(device)
+        if 'attention_mask' not in inputs:
+            inputs['attention_mask'] = text_inputs['attention_mask'].to(device)
+
+        # 尝试添加 decoder_input_ids
+        bos_token_id = processor.tokenizer.bos_token_id # 获取 BOS token ID
+        if bos_token_id is None: # 如果 BOS token ID 为 None，则使用 EOS token ID 作为 fallback
+            bos_token_id = processor.tokenizer.eos_token_id
+        decoder_input_ids = torch.tensor([[bos_token_id] * len(batch)]).T.to(device) # 创建 decoder_input_ids，形状为 (batch_size, 1)
+        inputs['decoder_input_ids'] = decoder_input_ids # 将 decoder_input_ids 添加到 inputs 字典
+
+        print("Before model.generate call:")
+        print(f"  Input attention_mask shape: {inputs['attention_mask'].shape if 'attention_mask' in inputs else 'No attention_mask in inputs'}")
+        print(f"  Input pixel_values shape: {inputs['pixel_values'].shape if 'pixel_values' in inputs else 'No pixel_values in inputs'}")
+        print(f"  Input input_ids shape: {inputs['input_ids'].shape if 'input_ids' in inputs else 'No input_ids in inputs'}")
+        print(f"  Full inputs dictionary: {inputs}")
+        print(f"  Inputs dictionary keys: {inputs.keys()}") # 打印 inputs 字典的键
+
+        if 'florence' in model.config.name_or_path:
+            generated_ids = model.generate(**inputs, max_length=100, num_beams=5, no_repeat_ngram_size=2, early_stopping=True, num_return_sequences=1) # temperature=0.01, do_sample=True,
+        elif 'blip2' in model.config.name_or_path:
+            generated_ids = model.generate(**inputs, max_new_tokens=20)
+        elif 'phi3_v' in model.config.model_type:
+            generated_ids = model.generate(**inputs, max_new_tokens=50)
+        else: # clip-tag and mocov3
+            generated_ids = model(**inputs)
+
+        print("After model.generate call:") # 添加提示信息
+        if isinstance(generated_ids, tuple): # 处理不同模型输出
+            generated_ids = generated_ids[0] # 假设第一个元素是 generated_ids
+        print(f"  Generated IDs object: {generated_ids}") # 打印 generated_ids 对象
+        print(f"  Generated IDs shape: {generated_ids.sequences.shape}") # 修改为访问 sequences 属性
+        generated_ids_for_decode = generated_ids.sequences #  获取 sequences 属性用于解码
+
+        generated_texts_batch = processor.batch_decode(generated_ids_for_decode, skip_special_tokens=True) # 使用 sequences 属性进行解码
+        generated_texts.extend(generated_texts_batch)
+        end = time.time()
+        print(f"batch {i//batch_size} takes {end-start:.2f} seconds, infer time {end-t1:.2f} seconds, batch size {len(batch)}, generated text: {generated_texts_batch}")
+
+    parsed_content_icon = generated_texts # 先进行赋值
+    return parsed_content_icon          # 然后返回 parsed_content_icon
+
+
+
+def get_parsed_content_icon_phi3v(filtered_boxes, ocr_bbox, image_source, caption_model_processor):
+    to_pil = ToPILImage()
+    if ocr_bbox:
+        non_ocr_boxes = filtered_boxes[len(ocr_bbox):]
+    else:
+        non_ocr_boxes = filtered_boxes
+    croped_pil_image = []
+    for i, coord in enumerate(non_ocr_boxes):
+        xmin, xmax = int(coord[0]*image_source.shape[1]), int(coord[2]*image_source.shape[1])
+        ymin, ymax = int(coord[1]*image_source.shape[0]), int(coord[3]*image_source.shape[0])
+        cropped_image = image_source[ymin:ymax, xmin:xmax, :]
+        croped_pil_image.append(to_pil(cropped_image))
+
+    model, processor = caption_model_processor['model'], caption_model_processor['processor']
+    device = model.device
+    messages = [{"role": "user", "content": "<|image_1|>\ndescribe the icon in one sentence"}] 
+    prompt = processor.tokenizer.apply_chat_template(messages, tokenize=False, add_generation_prompt=True)
+
+    batch_size = 5  # Number of samples per batch
+    generated_texts = []
+
+    for i in range(0, len(croped_pil_image), batch_size):
+        images = croped_pil_image[i:i+batch_size]
+        image_inputs = [processor.image_processor(x, return_tensors="pt") for x in images]
+        inputs ={'input_ids': [], 'attention_mask': [], 'pixel_values': [], 'image_sizes': []}
+        texts = [prompt] * len(images)
+        for i, txt in enumerate(texts):
+            input = processor._convert_images_texts_to_inputs(image_inputs[i], txt, return_tensors="pt")
+            inputs['input_ids'].append(input['input_ids'])
+            inputs['attention_mask'].append(input['attention_mask'])
+            inputs['pixel_values'].append(input['pixel_values'])
+            inputs['image_sizes'].append(input['image_sizes'])
+        max_len = max([x.shape[1] for x in inputs['input_ids']])
+        for i, v in enumerate(inputs['input_ids']):
+            inputs['input_ids'][i] = torch.cat([processor.tokenizer.pad_token_id * torch.ones(1, max_len - v.shape[1], dtype=torch.long), v], dim=1)
+            inputs['attention_mask'][i] = torch.cat([torch.zeros(1, max_len - v.shape[1], dtype=torch.long), inputs['attention_mask'][i]], dim=1)
+        inputs_cat = {k: torch.concatenate(v).to(device) for k, v in inputs.items()}
+
+        generation_args = { 
+            "max_new_tokens": 25, 
+            "temperature": 0.01, 
+            "do_sample": False, 
+        } 
+        generate_ids = model.generate(**inputs_cat, eos_token_id=processor.tokenizer.eos_token_id, **generation_args) 
+        # # remove input tokens 
+        generate_ids = generate_ids[:, inputs_cat['input_ids'].shape[1]:]
+        response = processor.batch_decode(generate_ids, skip_special_tokens=True, clean_up_tokenization_spaces=False)
+        response = [res.strip('\n').strip() for res in response]
+        generated_texts.extend(response)
+
+    return generated_texts
+
+def remove_overlap(boxes, iou_threshold, ocr_bbox=None):
+    assert ocr_bbox is None or isinstance(ocr_bbox, List)
+
+    def box_area(box):
+        return (box[2] - box[0]) * (box[3] - box[1])
+
+    def intersection_area(box1, box2):
+        x1 = max(box1[0], box2[0])
+        y1 = max(box1[1], box2[1])
+        x2 = min(box1[2], box2[2])
+        y2 = min(box1[3], box2[3])
+        return max(0, x2 - x1) * max(0, y2 - y1)
+
+    def IoU(box1, box2):
+        intersection = intersection_area(box1, box2)
+        union = box_area(box1) + box_area(box2) - intersection + 1e-6
+        if box_area(box1) > 0 and box_area(box2) > 0:
+            ratio1 = intersection / box_area(box1)
+            ratio2 = intersection / box_area(box2)
+        else:
+            ratio1, ratio2 = 0, 0
+        return max(intersection / union, ratio1, ratio2)
+
+    def is_inside(box1, box2):
+        # return box1[0] >= box2[0] and box1[1] >= box2[1] and box1[2] <= box2[2] and box1[3] <= box2[3]
+        intersection = intersection_area(box1, box2)
+        ratio1 = intersection / box_area(box1)
+        return ratio1 > 0.95
+
+    boxes = boxes.tolist()
+    filtered_boxes = []
+    if ocr_bbox:
+        filtered_boxes.extend(ocr_bbox)
+    # print('ocr_bbox!!!', ocr_bbox)
+    for i, box1 in enumerate(boxes):
+        # if not any(IoU(box1, box2) > iou_threshold and box_area(box1) > box_area(box2) for j, box2 in enumerate(boxes) if i != j):
+        is_valid_box = True
+        for j, box2 in enumerate(boxes):
+            # keep the smaller box
+            if i != j and IoU(box1, box2) > iou_threshold and box_area(box1) > box_area(box2):
+                is_valid_box = False
+                break
+        if is_valid_box:
+            # add the following 2 lines to include ocr bbox
+            if ocr_bbox:
+                # only add the box if it does not overlap with any ocr bbox
+                if not any(IoU(box1, box3) > iou_threshold and not is_inside(box1, box3) for k, box3 in enumerate(ocr_bbox)):
+                    filtered_boxes.append(box1)
+            else:
+                filtered_boxes.append(box1)
+    return torch.tensor(filtered_boxes)
+
+def remove_overlap_new(boxes, iou_threshold, ocr_bbox=None):
+    if not ocr_bbox:
+        return boxes
+
+    def box_area(box):
+        return (box[2] - box[0]) * (box[3] - box[1])
+
+    def intersection_area(box1, box2):
+        x1 = max(box1[0], box2[0])
+        y1 = max(box1[1], box2[1])
+        x2 = min(box1[2], box2[2])
+        y2 = min(box1[3], box2[3])
+        return max(0, x2 - x1) * max(0, y2 - y1)
+
+    def IoU(box1, box2):
+        intersection = intersection_area(box1, box2)
+        union = box_area(box1) + box_area(box2) - intersection + 1e-6
+        if box_area(box1) > 0 and box_area(box2) > 0:
+            ratio1 = intersection / box_area(box1)
+            ratio2 = intersection / box_area(box2)
+        else:
+            ratio1, ratio2 = 0, 0
+        return max(intersection / union, ratio1, ratio2)
+
+    def is_inside(box1, box2):
+        intersection = intersection_area(box1, box2)
+        ratio1 = intersection / box_area(box1)
+        return ratio1 > 0.80  # 可调整阈值
+
+    filtered_boxes = []
+    if ocr_bbox:
+        filtered_boxes.extend(ocr_bbox)
+
+    for i, box1_elem in enumerate(boxes):
+        box1 = box1_elem['bbox']
+        is_valid_box = True
+        for j, box2_elem in enumerate(boxes):
+            box2 = box2_elem['bbox']
+            if i != j and IoU(box1, box2) > iou_threshold and box_area(box1) > box_area(box2):
+                is_valid_box = False
+                break
+
+        if is_valid_box:
+            if ocr_bbox:
+                box_added = False
+                ocr_labels = ''
+                indices_to_remove = []  # 存储需要移除的 ocr_bbox 元素的索引
+                for k, box3_elem in enumerate(ocr_bbox):
+                    if not box_added:
+                        box3 = box3_elem['bbox']
+                        if is_inside(box3, box1):
+                            try:
+                                ocr_labels += box3_elem['content'] + ' '
+                                indices_to_remove.append(k)  # 记录索引
+                            except:
+                                continue
+                        elif is_inside(box1, box3):
+                            box_added = True
+                            break
+                        else:
+                            continue
+
+                # 逆序移除元素，避免索引错位
+                for index in sorted(indices_to_remove, reverse=True):
+                    filtered_boxes.pop(index)
+
+                if not box_added:
+                    if ocr_labels:
+                        filtered_boxes.append({'type': 'icon', 'bbox': box1_elem['bbox'], 'interactivity': True, 'content': ocr_labels, 'source':'box_yolo_content_ocr'})
+                    else:
+                        filtered_boxes.append({'type': 'icon', 'bbox': box1_elem['bbox'], 'interactivity': True, 'content': None, 'source':'box_yolo_content_yolo'})
+            else:
+                filtered_boxes.append(box1_elem) # 修改：添加box1_elem而不是box1
+    return filtered_boxes
+
+def load_image(image_path: str) -> Tuple[np.array, torch.Tensor]:
+    transform = T.Compose(
+        [
+            T.RandomResize([800], max_size=1333),
+            T.ToTensor(),
+            T.Normalize([0.485, 0.456, 0.406], [0.229, 0.224, 0.225]),
+        ]
+    )
+    image_source = Image.open(image_path).convert("RGB")
+    image = np.asarray(image_source)
+    image_transformed, _ = transform(image_source, None)
+    return image, image_transformed
+
+
+def annotate(image_source: np.ndarray, boxes: torch.Tensor, logits: torch.Tensor, phrases: List[str], text_scale: float, 
+             text_padding=5, text_thickness=2, thickness=3) -> np.ndarray:
+    """    
+    This function annotates an image with bounding boxes and labels.
+
+    Parameters:
+    image_source (np.ndarray): The source image to be annotated.
+    boxes (torch.Tensor): A tensor containing bounding box coordinates. in cxcywh format, pixel scale
+    logits (torch.Tensor): A tensor containing confidence scores for each bounding box.
+    phrases (List[str]): A list of labels for each bounding box.
+    text_scale (float): The scale of the text to be displayed. 0.8 for mobile/web, 0.3 for desktop # 0.4 for mind2web
+
+    Returns:
+    np.ndarray: The annotated image.
+    """
+    h, w, _ = image_source.shape
+    boxes = boxes * torch.Tensor([w, h, w, h])
+    xyxy = box_convert(boxes=boxes, in_fmt="cxcywh", out_fmt="xyxy").numpy()
+    xywh = box_convert(boxes=boxes, in_fmt="cxcywh", out_fmt="xywh").numpy()
+    detections = sv.Detections(xyxy=xyxy)
+
+    labels = [f"{phrase}" for phrase in range(boxes.shape[0])]
+
+    box_annotator = BoxAnnotator(text_scale=text_scale, text_padding=text_padding,text_thickness=text_thickness,thickness=thickness) # 0.8 for mobile/web, 0.3 for desktop # 0.4 for mind2web
+    annotated_frame = image_source.copy()
+    annotated_frame = box_annotator.annotate(scene=annotated_frame, detections=detections, labels=labels, image_size=(w,h))
+
+    label_coordinates = {f"{phrase}": v for phrase, v in zip(phrases, xywh)}
+    return annotated_frame, label_coordinates
+
+
+def predict(model, image, caption, box_threshold, text_threshold):
+    """ Use huggingface model to replace the original model
+    """
+    model, processor = model['model'], model['processor']
+    device = model.device
+
+    inputs = processor(images=image, text=caption, return_tensors="pt").to(device)
+    with torch.no_grad():
+        outputs = model(**inputs)
+
+    results = processor.post_process_grounded_object_detection(
+        outputs,
+        inputs.input_ids,
+        box_threshold=box_threshold, # 0.4,
+        text_threshold=text_threshold, # 0.3,
+        target_sizes=[image.size[::-1]]
+    )[0]
+    boxes, logits, phrases = results["boxes"], results["scores"], results["labels"]
+    return boxes, logits, phrases
+
+
+def predict_yolo(model, image, box_threshold, imgsz, scale_img, iou_threshold=0.7):
+    """ Use huggingface model to replace the original model
+    """
+    # model = model['model']
+    if scale_img:
+        result = model.predict(
+        source=image,
+        conf=box_threshold,
+        imgsz=imgsz,
+        iou=iou_threshold, # default 0.7
+        )
+    else:
+        result = model.predict(
+        source=image,
+        conf=box_threshold,
+        iou=iou_threshold, # default 0.7
+        )
+    boxes = result[0].boxes.xyxy#.tolist() # in pixel space
+    conf = result[0].boxes.conf
+    phrases = [str(i) for i in range(len(boxes))]
+
+    return boxes, conf, phrases
+
+def int_box_area(box, w, h):
+    x1, y1, x2, y2 = box
+    int_box = [int(x1*w), int(y1*h), int(x2*w), int(y2*h)]
+    area = (int_box[2] - int_box[0]) * (int_box[3] - int_box[1])
+    return area
+
+def get_som_labeled_img(image_source: Union[str, Image.Image], model=None, BOX_TRESHOLD=0.01, output_coord_in_ratio=False, ocr_bbox=None, text_scale=0.4, text_padding=5, draw_bbox_config=None, caption_model_processor=None, ocr_text=[], use_local_semantics=True, iou_threshold=0.9,prompt=None, scale_img=False, imgsz=None, batch_size=128):
+    """Process either an image path or Image object
+    
+    Args:
+        image_source: Either a file path (str) or PIL Image object
+        ...
+    """
+    
+    print(f"get_som_labeled_img 开始：ocr_bbox = {ocr_bbox}")
+    
+    print(f"ocr_text 类型：{type(ocr_text)}")
+    print(f"ocr_text 内容：{ocr_text}")
+    (ocr_text, ocr_bbox) = check_ocr_box(image_source, display_img=False, output_bb_format='xywh', goal_filtering=None, easyocr_args=None, use_paddleocr=False)
+    
+     # 检查 ocr_bbox 是否为空，如果为空，则初始化为空列表
+    if ocr_bbox is None:
+        ocr_bbox = []
+        
+    if isinstance(image_source, str):
+        image_source = Image.open(image_source)
+    image_source = image_source.convert("RGB")  # for CLIP
+    w, h = image_source.size
+    if not imgsz:
+        imgsz = (h, w)
+    # print('image size:', w, h)
+    xyxy, logits, phrases = predict_yolo(model=model, image=image_source, box_threshold=BOX_TRESHOLD, imgsz=imgsz, scale_img=scale_img, iou_threshold=0.1)
+    xyxy = xyxy / torch.Tensor([w, h, w, h]).to(xyxy.device)
+    image_source = np.asarray(image_source)
+    phrases = [str(i) for i in range(len(phrases))]
+
+    # annotate the image with labels
+    # 正确处理空列表和 None 值
+    if ocr_bbox is not None and len(ocr_bbox) > 0:  # 修改后的条件判断
+        print("准备转换 ocr_bbox 类型")  # 在类型转换之前
+        ocr_bbox = torch.tensor(ocr_bbox) / torch.Tensor([w, h, w, h])
+        print(f"转换 ocr_bbox 类型后：ocr_bbox = {ocr_bbox}")
+
+        print("准备将 ocr_bbox 转换为列表")  # 在转换为列表之前
+        ocr_bbox = ocr_bbox.tolist()
+        print(f"将 ocr_bbox 转换为列表后：ocr_bbox = {ocr_bbox}")
+    else:
+        print('no ocr bbox!!!')
+        ocr_bbox = []  # 赋值为空列表，而不是 None
+
+    print(f"get_som_labeled_img AFTER OCR BBOX PROCESSING: ocr_bbox = {ocr_bbox}")  # 2. Print after processing ocr_bbox
+
+    print(f"get_som_labeled_img AFTER OCR BBOX PROCESSING: ocr_bbox = {ocr_bbox}") # 2. Print after processing ocr_bbox
+
+    print("About to create ocr_bbox_elem") # 3. Before ocr_bbox_elem creation
+     # 检查 ocr_bbox 和 ocr_text 是否都为空
+    if ocr_bbox and ocr_text: # 两个列表都不为空时才执行
+        ocr_bbox_elem = [{'type': 'text', 'bbox':box, 'interactivity':False, 'content':txt, 'source': 'box_ocr_content_ocr'} for box, txt in zip(ocr_bbox, ocr_text) if int_box_area(box, w, h) > 0]
+    else:
+        ocr_bbox_elem = [] # 赋予空列表
+    
+    print(f"ocr_bbox_elem created: ocr_bbox = {ocr_bbox}")
+
+    print(f"ocr_bbox_elem created: ocr_bbox = {ocr_bbox}") # 4. After ocr_bbox_elem creation
+
+    print("About to create xyxy_elem") # 5. Before xyxy_elem creation
+    xyxy_elem = [{'type': 'icon', 'bbox': box, 'interactivity': True, 'content': None} for box in xyxy.tolist() if int_box_area(box, w, h) > 0]
+    print(f"xyxy_elem created: ocr_bbox = {ocr_bbox}") # 6. After xyxy_elem creation
+
+    print("About to create filtered_boxes") # 7. Before filtered_boxes creation
+    filtered_boxes = remove_overlap_new(boxes=xyxy_elem, iou_threshold=iou_threshold, ocr_bbox=ocr_bbox_elem)
+    print(f"filtered_boxes created: ocr_bbox = {ocr_bbox}") # 8. After filtered_boxes creation
+
+    print(f"ocr_text 类型：{type(ocr_text)}")
+    print(f"ocr_text 内容：{ocr_text}")
+
+     # 检查 ocr_bbox 和 ocr_text 是否都为空
+    if ocr_bbox and ocr_text: # 两个列表都不为空时才执行
+        ocr_bbox_elem = [{'type': 'text', 'bbox':box, 'interactivity':False, 'content':txt, 'source': 'box_ocr_content_ocr'} for box, txt in zip(ocr_bbox, ocr_text) if int_box_area(box, w, h) > 0]
+    else:
+        ocr_bbox_elem = [] # 赋予空列表
+        
+
+    xyxy_elem = [{'type': 'icon', 'bbox':box, 'interactivity':True, 'content':None} for box in xyxy.tolist() if int_box_area(box, w, h) > 0]
+    filtered_boxes = remove_overlap_new(boxes=xyxy_elem, iou_threshold=iou_threshold, ocr_bbox=ocr_bbox_elem)
+
+    # sort the filtered_boxes so that the one with 'content': None is at the end, and get the index of the first 'content': None
+    filtered_boxes_elem = sorted(filtered_boxes, key=lambda x: x['content'] is None)
+    # get the index of the first 'content': None
+    starting_idx = next((i for i, box in enumerate(filtered_boxes_elem) if box['content'] is None), -1)
+    filtered_boxes = torch.tensor([box['bbox'] for box in filtered_boxes_elem])
+    print('len(filtered_boxes):', len(filtered_boxes), starting_idx)
+
+    # get parsed icon local semantics
+    time1 = time.time()
+    caption_model = caption_model_processor['model'] if caption_model_processor else None
+
+    if caption_model_processor is None: # 增加 caption_model_processor 的 None 值检查
+        print("警告: caption_model_processor 为 None，图像描述功能可能无法使用。")
+    elif caption_model is not None: # 保留原有的 caption_model 的 None 值检查
+        if 'phi3_v' in caption_model.config.model_type:
+            parsed_content_icon = get_parsed_content_icon_phi3v(filtered_boxes, ocr_bbox, image_source, caption_model_processor)
+        else:
+            print("Before get_parsed_content_icon call:") # 添加提示信息
+            print(f"  filtered_boxes shape: {filtered_boxes.shape}") # 打印 filtered_boxes 的形状
+            print(f"  image_source type: {type(image_source)}") # 打印 image_source 的类型
+            print(f"  caption_model_processor: {caption_model_processor}") # 打印 caption_model_processor
+            parsed_content_icon = get_parsed_content_icon(filtered_boxes, starting_idx, image_source, caption_model_processor, prompt=prompt,batch_size=batch_size)
+        ocr_text = [f"Text Box ID {i}: {txt}" for i, txt in enumerate(ocr_text)]
+        icon_start = len(ocr_text)
+        parsed_content_icon_ls = []
+        # fill the filtered_boxes_elem None content with parsed_content_icon in order
+        for i, box in enumerate(filtered_boxes_elem):
+            if box['content'] is None:
+                box['content'] = parsed_content_icon.pop(0)
+        for i, txt in enumerate(parsed_content_icon):
+            parsed_content_icon_ls.append(f"Icon Box ID {str(i+icon_start)}: {txt}")
+        parsed_content_merged = ocr_text + parsed_content_icon_ls
+    else:
+        ocr_text = [f"Text Box ID {i}: {txt}" for i, txt in enumerate(ocr_text)]
+        parsed_content_merged = ocr_text
+    print('time to get parsed content:', time.time()-time1)
+
+    filtered_boxes = box_convert(boxes=filtered_boxes, in_fmt="xyxy", out_fmt="cxcywh")
+
+    phrases = [i for i in range(len(filtered_boxes))]
+    
+    # draw boxes
+    if draw_bbox_config:
+        annotated_frame, label_coordinates = annotate(image_source=image_source, boxes=filtered_boxes, logits=logits, phrases=phrases, **draw_bbox_config)
+    else:
+        annotated_frame, label_coordinates = annotate(image_source=image_source, boxes=filtered_boxes, logits=logits, phrases=phrases, text_scale=text_scale, text_padding=text_padding)
+    
+    pil_img = Image.fromarray(annotated_frame)
+    buffered = io.BytesIO()
+    pil_img.save(buffered, format="PNG")
+    encoded_image = base64.b64encode(buffered.getvalue()).decode('ascii')
+    if output_coord_in_ratio:
+        label_coordinates = {k: [v[0]/w, v[1]/h, v[2]/w, v[3]/h] for k, v in label_coordinates.items()}
+        assert w == annotated_frame.shape[1] and h == annotated_frame.shape[0]
+
+    return encoded_image, label_coordinates, filtered_boxes_elem
+
+
+def get_xywh(input):
+    x, y, w, h = input[0][0], input[0][1], input[2][0] - input[0][0], input[2][1] - input[0][1]
+    x, y, w, h = int(x), int(y), int(w), int(h)
+    return x, y, w, h
+
+def get_xyxy(input):
+    x, y, xp, yp = input[0][0], input[0][1], input[2][0], input[2][1]
+    x, y, xp, yp = int(x), int(y), int(xp), int(yp)
+    return x, y, xp, yp
+
+def get_xywh_yolo(input):
+    x, y, w, h = input[0], input[1], input[2] - input[0], input[3] - input[1]
+    x, y, w, h = int(x), int(y), int(w), int(h)
+    return x, y, w, h
+
+def check_ocr_box(image_source: Union[str, Image.Image], display_img = True, output_bb_format='xywh', goal_filtering=None, easyocr_args=None, use_paddleocr=False):
+    if image_source is ...:  # 检查 image_source 是否为 ...
+        print("错误：image_source 是一个 ellipsis 对象！")
+        return ([], [])  # 或者其他适当的处理方式
+
+    if isinstance(image_source, str):
+        image_source = Image.open(image_source)
+    if isinstance(image_source, str):
+        image_source = Image.open(image_source)
+    if image_source.mode == 'RGBA':
+        # Convert RGBA to RGB to avoid alpha channel issues
+        image_source = image_source.convert('RGB')
+    image_np = np.array(image_source)
+    w, h = image_source.size
+    if use_paddleocr:
+        if easyocr_args is None:
+            text_threshold = 0.5
+        else:
+            text_threshold = easyocr_args['text_threshold']
+        result = paddle_ocr.ocr(image_np, cls=False)[0]
+        coord = [item[0] for item in result if item[1][1] > text_threshold]
+        text = [item[1][0] for item in result if item[1][1] > text_threshold]
+    else:  # EasyOCR
+        if easyocr_args is None:
+            easyocr_args = {}
+        result = reader.readtext(image_np, **easyocr_args)
+        coord = [item[0] for item in result]
+        text = [item[1] for item in result]
+    if display_img:
+        opencv_img = cv2.cvtColor(image_np, cv2.COLOR_RGB2BGR)
+        bb = []
+        for item in coord:
+            x, y, a, b = get_xywh(item)
+            bb.append((x, y, a, b))
+            cv2.rectangle(opencv_img, (x, y), (x+a, y+b), (0, 255, 0), 2)
+        #  matplotlib expects RGB
+        plt.imshow(cv2.cvtColor(opencv_img, cv2.COLOR_BGR2RGB))
+    else:
+        if output_bb_format == 'xywh':
+            bb = [get_xywh(item) for item in coord]
+        elif output_bb_format == 'xyxy':
+            bb = [get_xyxy(item) for item in coord]
+    return (text, bb), goal_filtering