# ---------------------------------------------------------------------- # IMPORTS # ---------------------------------------------------------------------- import torch import numpy as np import logging import time import spaces import sys from PIL import Image, ImageDraw from collections import defaultdict from typing import Dict, List, Optional, Tuple, Any, Union from src.utils import ProcessingContext, create_pipeline_step, LOG_LEVEL_MAP, EMOJI_MAP from src.processing.bounding_box.yolos_fashionpedia_model import ( detect_with_yolos_fashionpedia, detect_yolos_in_roi, process_yolos_results ) from src.processing.bounding_box.rtdetr_model import ( detect_rtdetr_in_roi, detect_rtdetr_artifacts_in_roi, update_fallback_detection, get_rtdetr_clothing_labels, get_rtdetr_person_and_product_labels, get_rtdetr_artifact_labels, process_rtdetr_results ) from src.processing.bounding_box.head_model import detect_head_in_roi # ---------------------------------------------------------------------- # CONSTANTS # ---------------------------------------------------------------------- Apply_Draw = False DETECT_ARTIFACTS = False PRODUCT_TYPE_BOX_FILTER_THRESHOLD = 0.4 HEAD_BOX_FILTER_THRESHOLD = 0.5 SHOES_BOX_FILTER_THRESHOLD = 0.5 CLOTHING_FEATURES_BOX_FILTER_THRESHOLD = 0.4 ARTIFACTS_BOX_FILTER_THRESHOLD = 0.7 BOX_OVERLAP = 50 LARGEST_BOX_EXTENSION_RATIO = 0.02 LARGEST_BOX_THRESHOLD = 5 CREATE_UPPER_BLUE_LOWER_GREEN_RATIO = 0.02 CREATE_LOWER_BLUE_LOWER_VIOLET_RATIO = 0.5 LEFT_RIGHT_BLUE_OFFSET = 0.5 LOWER_UPPER_BLUE_OFFSET = 0.02 BLUE_BOX_FALLBACK_THRESHOLD = 33 UPPER_HEAD_FILTER = 33 LOWER_SHOE_FILTER = 33 NECKLINE_COLLAR_HEAD = 50 BLUE_BOX_PRODUCT_TYPE = (0, 0, 255, 200) GREEN_BOX_HEAD = (0, 255, 0, 200) VIOLET_BOX_SHOES = (238, 130, 238, 200) ORANGE_BOX_CLOTHING_FEATURES = (255, 165, 0, 200) RED_BOX_ARTIFACTS = (255, 0, 0, 200) BLACK_BOX_PERSON = (0, 0, 0, 200) # ---------------------------------------------------------------------- # DYNAMIC LIST EXTRACTION # ---------------------------------------------------------------------- def extract_category_lists(): from .yolos_fashionpedia_model import MODEL_LABEL_CONFIG as YOLOS_CONFIG from .rtdetr_model import MODEL_LABEL_CONFIG as RTDETR_CONFIG from .head_model import MODEL_LABEL_CONFIG as HEAD_CONFIG MODEL_LABEL_CONFIG = {**YOLOS_CONFIG, **RTDETR_CONFIG, **HEAD_CONFIG} lists = { "PERSON_LIST": set(), "PRODUCT_TYPE_LIST": set(), "HEAD_LIST": set(), "SHOES_LIST": set(), "CLOTHING_FEATURES_LIST": set(), "ARTIFACTS_LIST": set() } for model_name, model_config in MODEL_LABEL_CONFIG.items(): for list_type, keywords in model_config.items(): list_key = list_type.upper() if list_key in lists: lists[list_key].update(keywords.keys()) return {k: list(v) for k, v in lists.items()} extracted_lists = extract_category_lists() PERSON_LIST = extracted_lists["PERSON_LIST"] PRODUCT_TYPE_LIST = extracted_lists["PRODUCT_TYPE_LIST"] HEAD_LIST = extracted_lists["HEAD_LIST"] SHOES_LIST = extracted_lists["SHOES_LIST"] CLOTHING_FEATURES_LIST = extracted_lists["CLOTHING_FEATURES_LIST"] ARTIFACTS_LIST = extracted_lists["ARTIFACTS_LIST"] # ---------------------------------------------------------------------- # SHARED UTILITY FUNCTIONS # ---------------------------------------------------------------------- def get_label_name_from_model(model, label_id): if hasattr(model, 'config') and hasattr(model.config, 'id2label'): return model.config.id2label.get(label_id, f"unknown_{label_id}").lower() if hasattr(model, 'model_labels') and isinstance(model.model_labels, dict): return model.model_labels.get(label_id, f"unknown_{label_id}").lower() return f"unknown_{label_id}" def clamp_box_to_region(box: List[int], region: List[int]) -> List[int]: x1, y1, x2, y2 = box rx1, ry1, rx2, ry2 = region xx1 = max(rx1, min(x1, rx2)) yy1 = max(ry1, min(y1, ry2)) xx2 = max(rx1, min(x2, rx2)) yy2 = max(ry1, min(y2, ry2)) return [xx1, yy1, xx2, yy2] def box_iou(b1, b2): xx1 = max(b1[0], b2[0]) yy1 = max(b1[1], b2[1]) xx2 = min(b1[2], b2[2]) yy2 = min(b1[3], b2[3]) iw = max(0, xx2 - xx1) ih = max(0, yy2 - yy1) inter_area = iw * ih area1 = abs(b1[2] - b1[0]) * abs(b1[3] - b1[1]) area2 = abs(b2[2] - b2[0]) * abs(b2[3] - b2[1]) union = area1 + area2 - inter_area if union <= 0: return 0 return inter_area / union def build_keywords(product_type: str) -> List[str]: pt = product_type.lower().strip() if not pt or pt == "unknown": return [] kw = [pt] if pt in ["jeans", "shorts", "skirt"]: kw += ["shoes", "closure", "pocket"] elif pt in ["jacket", "vest", "shirt"]: kw += ["head", "closure", "pocket", "collar", "sleeve"] elif pt in ["overall", "dress"]: kw += ["head", "shoes", "closure", "pocket", "neckline", "sleeve"] if DETECT_ARTIFACTS: kw += ["bag", "cup"] return kw def apply_box_overlap_filter( boxes: List[List[int]], labels: List[int], scores: List[float], keywords: List[str], raw_labels: List[str], models: List[str] ): grouped = defaultdict(list) for i in range(len(boxes)): kwd = keywords[i] grouped[kwd].append({ "box": boxes[i], "label": labels[i], "score": scores[i], "raw_label": raw_labels[i], "model": models[i] }) new_boxes = [] new_labels = [] new_scores = [] new_kws = [] new_raws = [] new_models = [] overlap_thresh = BOX_OVERLAP / 100.0 for kwd, items in grouped.items(): sorted_items = sorted(items, key=lambda x: x["score"], reverse=True) if sorted_items: best_item = sorted_items[0] new_boxes.append(best_item["box"]) new_labels.append(best_item["label"]) new_scores.append(round(best_item["score"], 2)) new_kws.append(kwd) new_raws.append(best_item["raw_label"]) new_models.append(best_item["model"]) return new_boxes, new_labels, new_scores, new_kws, new_raws, new_models def get_threshold_for_keyword(kw: str) -> float: if kw in PRODUCT_TYPE_LIST: return PRODUCT_TYPE_BOX_FILTER_THRESHOLD elif kw in HEAD_LIST: return HEAD_BOX_FILTER_THRESHOLD elif kw in SHOES_LIST: return SHOES_BOX_FILTER_THRESHOLD elif kw in CLOTHING_FEATURES_LIST: return CLOTHING_FEATURES_BOX_FILTER_THRESHOLD elif kw in ARTIFACTS_LIST: return ARTIFACTS_BOX_FILTER_THRESHOLD return 0.0 def map_label_to_keyword(label_name: str, valid_kws: List[str], model_name: str) -> Optional[str]: ln = label_name.strip().lower() if ln in valid_kws: return ln if model_name == "yolos_fashionpedia_model": from .yolos_fashionpedia_model import MODEL_LABEL_CONFIG elif model_name == "rtdetr_model": from .rtdetr_model import MODEL_LABEL_CONFIG elif model_name == "head_model": from .head_model import MODEL_LABEL_CONFIG else: return None model_config = MODEL_LABEL_CONFIG.get(model_name, {}) for list_type in ["person_list", "product_type_list", "head_list", "shoes_list", "clothing_features_list", "artifacts_list"]: category_config = model_config.get(list_type, {}) for keyword, labels in category_config.items(): if keyword in valid_kws: for label in labels: if ln == label.lower() or ln in label.lower(): return keyword return None # ---------------------------------------------------------------------- # MODEL VALIDATION # ---------------------------------------------------------------------- def validate_model_configurations(): from .yolos_fashionpedia_model import MODEL_LABEL_CONFIG as YOLOS_CONFIG from .rtdetr_model import MODEL_LABEL_CONFIG as RTDETR_CONFIG from .head_model import MODEL_LABEL_CONFIG as HEAD_CONFIG MODEL_LABEL_CONFIG = {**YOLOS_CONFIG, **RTDETR_CONFIG, **HEAD_CONFIG} logging.info("=" * 70) logging.info("āœ… MODEL CONFIGURATIONS VALIDATION") logging.info("=" * 70) for model_name, model_config in MODEL_LABEL_CONFIG.items(): logging.info(f"\nšŸ“Œ {model_name.upper()}") logging.info("-" * 50) total_prompts = 0 for category, category_config in model_config.items(): category_prompts = 0 for keyword, labels in category_config.items(): category_prompts += len(labels) total_prompts += len(labels) if category_config: logging.info(f" šŸ“‚ {category}: {len(category_config)} keywords, {category_prompts} prompts") logging.info(f" šŸŽÆ Total: {total_prompts} prompts") if model_name == "yolos_fashionpedia_model": logging.info(" āœØ YOLOS Fashionpedia specializes in fashion detection!") logging.info("\n" + "=" * 70) logging.info("šŸš€ FEATURE CAPABILITIES:") logging.info("-" * 50) logging.info(f" ā€¢ YOLOS Fashionpedia Detection: ENABLED") logging.info(f" ā€¢ Artifact Detection: {'ENABLED' if DETECT_ARTIFACTS else 'DISABLED'}") logging.info(f" ā€¢ Apply Draw (Visualization): {'ENABLED' if Apply_Draw else 'DISABLED'}") logging.info("=" * 70) # ---------------------------------------------------------------------- # DETECTION PIPELINE FUNCTIONS # ---------------------------------------------------------------------- def define_largest_box_batch(contexts, batch_logs, RTDETR_PROCESSOR, RTDETR_MODEL, RTDETR_FULL_PRECISION, DEVICE, MODELS_LOADED, LOAD_ERROR): from .rtdetr_model import RTDETR_CONF function_name = "define_largest_box_batch" start_time = time.perf_counter() logging.log(LOG_LEVEL_MAP["INFO"], f"{EMOJI_MAP['INFO']} Starting {function_name} for {len(contexts)} items") items_for_batch = [] valid_ctx_indices = [] batch_items = [] image_sizes = [] for i, ctx in enumerate(contexts): item_ = { "image_url": ctx.url, "data": {"largest_box": None}, "function": function_name } if ctx.skip_run or ctx.skip_processing: item_["status"] = "skipped" batch_logs.append(item_) continue if not MODELS_LOADED: import sys import traceback error_msg = LOAD_ERROR or "Models not loaded" error_trace = traceback.format_exc() logging.error(f"CRITICAL: Model not loaded in {function_name}: {error_msg}") logging.error(f"Traceback:\n{error_trace}") item_["status"] = "critical_error" item_["exception"] = error_msg item_["traceback"] = error_trace ctx.skip_run = True ctx.error = error_msg ctx.error_traceback = error_trace batch_logs.append(item_) logging.critical("Terminating due to model loading failure") sys.exit(1) if "original" not in ctx.pil_img: item_["status"] = "error" item_["exception"] = "No RBC 'original' found" ctx.skip_run = True batch_logs.append(item_) continue pi_rgba, _, _ = ctx.pil_img["original"] rgb_img = pi_rgba.convert("RGB") items_for_batch.append(rgb_img) image_sizes.append([rgb_img.height, rgb_img.width]) valid_ctx_indices.append(i) batch_items.append(item_) if not items_for_batch: processing_time = time.perf_counter() - start_time logging.log(LOG_LEVEL_MAP["INFO"], f"{EMOJI_MAP['INFO']} {function_name}: No valid items for batch processing in {processing_time:.3f}s") return batch_logs try: results = process_rtdetr_batch( items_for_batch, image_sizes, RTDETR_PROCESSOR, RTDETR_MODEL, DEVICE ) rtdetr_clothing_labels = get_rtdetr_clothing_labels() rtdetr_person_product_labels = get_rtdetr_person_and_product_labels() for b_idx, i_ in enumerate(valid_ctx_indices): ctx = contexts[i_] item_ = batch_items[b_idx] pi_rgb = items_for_batch[b_idx] W, H = pi_rgb.size detection_log = { "model": "RT-DETR", "largest_area": 0, "total_area": W * H, "extension_ratio": LARGEST_BOX_EXTENSION_RATIO, "threshold": LARGEST_BOX_THRESHOLD, "raw_detections": [] } if isinstance(results, list): if b_idx < len(results): result = results[b_idx] else: result = results[0] else: result = results largest_area = 0 main_box = None for score, label, box in zip(result["scores"], result["labels"], result["boxes"]): label_id = label.item() score_val = score.item() x1, y1, x2, y2 = [int(val) for val in box.tolist()] label_name = get_label_name_from_model(RTDETR_MODEL, label_id) if label_name in rtdetr_person_product_labels: logging.log(LOG_LEVEL_MAP["INFO"], f"{EMOJI_MAP['INFO']} RT-DETR detected: {label_name} at score {score_val:.3f} | box=[{x1},{y1},{x2},{y2}]") detection_log["raw_detections"].append({ "box": [x1, y1, x2, y2], "score": round(score_val, 3), "label": label_name, "is_clothing": label_name in rtdetr_clothing_labels }) if label_name in rtdetr_clothing_labels: area_ = (x2 - x1) * (y2 - y1) if area_ > largest_area: largest_area = area_ main_box = [x1, y1, x2, y2] if main_box is not None: ratio_off = int(min(W, H) * LARGEST_BOX_EXTENSION_RATIO) x1, y1, x2, y2 = main_box x1e = max(0, x1 - ratio_off) y1e = max(0, y1 - ratio_off) x2e = min(W, x2 + ratio_off) y2e = min(H, y2 + ratio_off) main_box = [x1e, y1e, x2e, y2e] total_area = W * H if largest_area < total_area * (LARGEST_BOX_THRESHOLD / 100.0): main_box = [0, 0, W, H] logging.log(LOG_LEVEL_MAP["INFO"], f"{EMOJI_MAP['INFO']} Fallback to full image: box=[0,0,{W},{H}] | {ctx.url}") else: logging.log(LOG_LEVEL_MAP["INFO"], f"{EMOJI_MAP['INFO']} Largest box detected: box=[{x1e},{y1e},{x2e},{y2e}] | area={largest_area} | {ctx.url}") else: main_box = [0, 0, W, H] logging.log(LOG_LEVEL_MAP["INFO"], f"{EMOJI_MAP['INFO']} No clothing detected, using full image: box=[0,0,{W},{H}] | {ctx.url}") ctx.define_result["largest_box"] = main_box item_["data"]["largest_box"] = main_box detection_log["largest_area"] = largest_area item_["data"]["largest_box_detection"] = detection_log item_["status"] = "ok" except Exception as e: import sys import traceback error_msg = f"{function_name} error: {str(e)}" error_trace = traceback.format_exc() logging.error(f"CRITICAL: {error_msg}") logging.error(f"Traceback:\n{error_trace}") for b_idx, i_ in enumerate(valid_ctx_indices): ctx = contexts[i_] item_ = batch_items[b_idx] item_["status"] = "critical_error" item_["exception"] = error_msg item_["traceback"] = error_trace ctx.skip_run = True ctx.error = str(e) ctx.error_traceback = error_trace logging.critical("Terminating due to bounding box processing failure") sys.exit(1) for item_ in batch_items: batch_logs.append(item_) processing_time = time.perf_counter() - start_time logging.log(LOG_LEVEL_MAP["SUCCESS"], f"{EMOJI_MAP['SUCCESS']} Completed {function_name} for {len(batch_items)} items in {processing_time:.3f}s") return batch_logs def process_rtdetr_batch(images, image_sizes, RTDETR_PROCESSOR, RTDETR_MODEL, DEVICE): from .rtdetr_model import RTDETR_CONF try: rtdetr_inputs = RTDETR_PROCESSOR(images=images, return_tensors="pt").to(DEVICE) with torch.no_grad(): rtdetr_outputs = RTDETR_MODEL(**rtdetr_inputs) rtdetr_results = RTDETR_PROCESSOR.post_process_object_detection( rtdetr_outputs, target_sizes=torch.tensor(image_sizes).to(DEVICE), threshold=RTDETR_CONF ) return rtdetr_results except Exception as e: logging.error(f"RT-DETR batch processing error: {str(e)}") return [] def process_single_image_detection(roi_rgb, keywords, rx1, ry1, rW, rH, detect_artifacts, detect_head, YOLOS_PROCESSOR, YOLOS_MODEL, RTDETR_PROCESSOR, RTDETR_MODEL, HEAD_PROCESSOR, HEAD_MODEL, HEAD_DETECTION_FULL_PRECISION, DEVICE): log_item = {"warnings": []} yolos_boxes, yolos_labels, yolos_scores, yolos_raws = detect_yolos_in_roi( roi_rgb, keywords, YOLOS_PROCESSOR, YOLOS_MODEL, DEVICE, log_item ) rtdetr_boxes, rtdetr_labels, rtdetr_scores, rtdetr_raws = detect_rtdetr_in_roi( roi_rgb, RTDETR_PROCESSOR, RTDETR_MODEL, DEVICE, log_item ) if detect_artifacts: artifact_boxes, artifact_labels, artifact_scores, artifact_raws = detect_rtdetr_artifacts_in_roi( roi_rgb, keywords, RTDETR_PROCESSOR, RTDETR_MODEL, DEVICE, log_item ) else: artifact_boxes, artifact_labels, artifact_scores, artifact_raws = [], [], [], [] if detect_head: head_boxes, head_labels, head_scores, head_raws = detect_head_in_roi( roi_rgb, rx1, ry1, rW, rH, HEAD_PROCESSOR, HEAD_MODEL, HEAD_DETECTION_FULL_PRECISION, DEVICE, log_item ) else: head_boxes, head_labels, head_scores, head_raws = [], [], [], [] return { 'yolos': (yolos_boxes, yolos_labels, yolos_scores, yolos_raws), 'rtdetr': (rtdetr_boxes, rtdetr_labels, rtdetr_scores, rtdetr_raws), 'artifacts': (artifact_boxes, artifact_labels, artifact_scores, artifact_raws), 'head': (head_boxes, head_labels, head_scores, head_raws), 'warnings': log_item.get("warnings", []) } def detect_batch(contexts, batch_logs, HEAD_PROCESSOR, HEAD_MODEL, HEAD_DETECTION_FULL_PRECISION, RTDETR_PROCESSOR, RTDETR_MODEL, RTDETR_FULL_PRECISION, YOLOS_PROCESSOR, YOLOS_MODEL, DEVICE, MODELS_LOADED, LOAD_ERROR): function_name = "detect_batch" start_time = time.perf_counter() logging.log(LOG_LEVEL_MAP["INFO"], f"{EMOJI_MAP['INFO']} Starting {function_name} for {len(contexts)} items") for ctx in contexts: log_item = { "image_url": ctx.url, "data": { "detection_result_log": {}, "product_type": ctx.product_type, "keywords": ctx.keywords }, "function": function_name } logging.log(LOG_LEVEL_MAP["INFO"], f"{EMOJI_MAP['INFO']} Processing image: {ctx.url} | Product Type: {ctx.product_type} | Keywords: {ctx.keywords}") if ctx.skip_run or ctx.skip_processing: log_item["status"] = "skipped" batch_logs.append(log_item) continue if "original" not in ctx.pil_img: log_item["status"] = "error" log_item["exception"] = "No RBC 'original'" ctx.skip_run = True batch_logs.append(log_item) continue pi_rgba, _, _ = ctx.pil_img["original"] W, H = pi_rgba.size largest_box = ctx.define_result.get("largest_box") if (not largest_box or not isinstance(largest_box, list) or len(largest_box) != 4): log_item["status"] = "no_detection" log_item["data"]["detection_result_log"] = "no_main_box_detected" batch_logs.append(log_item) continue rx1, ry1, rx2, ry2 = largest_box rW = rx2 - rx1 rH = ry2 - ry1 logging.log(LOG_LEVEL_MAP["INFO"], f"{EMOJI_MAP['INFO']} ROI box: [{rx1},{ry1},{rx2},{ry2}] | size: {rW}x{rH} | {ctx.url}") roi = pi_rgba.crop((rx1, ry1, rx2, ry2)) roi_rgb = roi.convert("RGB") all_boxes = [] all_labels = [] all_scores = [] all_raw = [] all_models = [] def box_to_global(x1, y1, x2, y2): gx1 = int(rx1 + x1) gy1 = int(ry1 + y1) gx2 = int(rx1 + x2) gy2 = int(ry1 + y2) return [gx1, gy1, gx2, gy2] detect_artifacts = DETECT_ARTIFACTS and any(kw in ctx.keywords for kw in ARTIFACTS_LIST) lower_body_types = ["jeans", "shorts", "skirt"] detect_head = ctx.product_type.lower() not in lower_body_types detection_start = time.perf_counter() detection_results = process_single_image_detection( roi_rgb, ctx.keywords, rx1, ry1, rW, rH, detect_artifacts, detect_head, YOLOS_PROCESSOR, YOLOS_MODEL, RTDETR_PROCESSOR, RTDETR_MODEL, HEAD_PROCESSOR, HEAD_MODEL, HEAD_DETECTION_FULL_PRECISION, DEVICE ) yolos_boxes, yolos_labels, yolos_scores, yolos_raws = detection_results['yolos'] rtdetr_boxes, rtdetr_labels, rtdetr_scores, rtdetr_raws = detection_results['rtdetr'] artifact_boxes, artifact_labels, artifact_scores, artifact_raws = detection_results['artifacts'] head_boxes, head_labels, head_scores, head_raws = detection_results['head'] if detection_results['warnings']: log_item["warnings"] = log_item.get("warnings", []) + detection_results['warnings'] detection_time = time.perf_counter() - detection_start logging.log(LOG_LEVEL_MAP["INFO"], f"{EMOJI_MAP['INFO']} Detection completed in {detection_time:.3f}s") for i in range(len(yolos_boxes)): x1, y1, x2, y2 = yolos_boxes[i] global_box = box_to_global(x1, y1, x2, y2) global_box = clamp_box_to_region(global_box, [rx1, ry1, rx2, ry2]) all_boxes.append(global_box) all_labels.append(40000 + yolos_labels[i]) all_scores.append(yolos_scores[i]) all_raw.append(yolos_raws[i]) all_models.append("yolos_fashionpedia_model") for i in range(len(rtdetr_boxes)): x1, y1, x2, y2 = rtdetr_boxes[i] global_box = box_to_global(x1, y1, x2, y2) global_box = clamp_box_to_region(global_box, [rx1, ry1, rx2, ry2]) all_boxes.append(global_box) all_labels.append(10000 + rtdetr_labels[i]) all_scores.append(rtdetr_scores[i]) all_raw.append(rtdetr_raws[i]) all_models.append("rtdetr_model") for i in range(len(artifact_boxes)): x1, y1, x2, y2 = artifact_boxes[i] global_box = box_to_global(x1, y1, x2, y2) global_box = clamp_box_to_region(global_box, [rx1, ry1, rx2, ry2]) all_boxes.append(global_box) all_labels.append(20000 + artifact_labels[i]) all_scores.append(artifact_scores[i]) all_raw.append(artifact_raws[i]) all_models.append("rtdetr_artifact") for i in range(len(head_boxes)): all_boxes.append(head_boxes[i]) all_labels.append(9999) all_scores.append(head_scores[i]) all_raw.append(head_raws[i]) all_models.append("head_model") logging.log(LOG_LEVEL_MAP["INFO"], f"{EMOJI_MAP['INFO']} Total detections before filtering: {len(all_boxes)} - YOLOS: {sum(1 for m in all_models if m == 'yolos_fashionpedia_model')}, RTDETR: {sum(1 for m in all_models if m == 'rtdetr_model')}, Head: {sum(1 for m in all_models if m == 'head_model')}") final_keywords = [None] * len(all_boxes) for i_ in range(len(all_boxes)): lb_ = all_labels[i_] model_name = all_models[i_] if lb_ == 9999: if "head" in ctx.keywords: final_keywords[i_] = "head" else: label_str = all_raw[i_] mapped = map_label_to_keyword(label_str, ctx.keywords, model_name) if mapped: final_keywords[i_] = mapped keep_mask = [True] * len(all_boxes) for i_ in range(len(all_boxes)): cat_ = final_keywords[i_] sc_ = all_scores[i_] if not cat_: keep_mask[i_] = False else: threshold_needed = get_threshold_for_keyword(cat_) if sc_ < threshold_needed: keep_mask[i_] = False fb, fl, fs, fk, fr, fm = [], [], [], [], [], [] for i_ in range(len(all_boxes)): if keep_mask[i_]: fb.append(all_boxes[i_]) fl.append(all_labels[i_]) fs.append(all_scores[i_]) fk.append(final_keywords[i_]) fr.append(all_raw[i_]) fm.append(all_models[i_]) fb, fl, fs, fk, fr, fm = apply_box_overlap_filter( fb, fl, fs, fk, fr, fm ) second_pass = [] for i_ in range(len(fb)): cat_ = fk[i_] (bx1, by1, bx2, by2) = fb[i_] centerY = 0.5 * (by1 + by2) if cat_ == "head": if centerY <= (H * (UPPER_HEAD_FILTER / 100.0)): second_pass.append(i_) elif cat_ == "shoes": if centerY >= (H * (1 - (LOWER_SHOE_FILTER / 100.0))): second_pass.append(i_) else: second_pass.append(i_) collar_idx = [i_ for i_, kw_ in enumerate(fk) if kw_ in ["collar", "neckline"]] collar_boxes = [fb[x] for x in collar_idx] def intersect_area(a_, b_): xx1 = max(a_[0], b_[0]) yy1 = max(a_[1], b_[1]) xx2 = min(a_[2], b_[2]) yy2 = min(a_[3], b_[3]) iw = max(0, xx2 - xx1) ih = max(0, yy2 - yy1) return iw * ih final_keep = [] for idx2 in second_pass: kw_ = fk[idx2] if kw_ == "head": hx1, hy1, hx2, hy2 = fb[idx2] hArea = (hx2 - hx1) * (hy2 - hy1) if hArea <= 0: continue remove = False for cb_ in collar_boxes: ia = intersect_area([hx1, hy1, hx2, hy2], cb_) if ia / float(hArea) > (NECKLINE_COLLAR_HEAD / 100.0): remove = True break if not remove: final_keep.append(idx2) else: final_keep.append(idx2) final_boxes = [fb[x] for x in final_keep] final_labels = [fl[x] for x in final_keep] final_scores = [round(fs[x], 2) for x in final_keep] final_kws = [fk[x] for x in final_keep] final_raws = [fr[x] for x in final_keep] final_mods = [fm[x] for x in final_keep] logging.log(LOG_LEVEL_MAP["INFO"], f"{EMOJI_MAP['INFO']} Final detections after filtering: {len(final_boxes)} boxes") for i in range(len(final_boxes)): box = final_boxes[i] logging.log(LOG_LEVEL_MAP["INFO"], f"{EMOJI_MAP['INFO']} Final detection: {final_kws[i]} ({final_raws[i]}) at score {final_scores[i]} | box=[{box[0]},{box[1]},{box[2]},{box[3]}]") dd_log = defaultdict(list) for i in range(len(final_boxes)): cat_ = final_kws[i] if cat_: dd_log[cat_].append({ "box": final_boxes[i], "score": final_scores[i], "raw_label": final_raws[i], "model": final_mods[i] }) if ctx.product_type not in dd_log: from .rtdetr_model import RTDETR_CONF fallback_box = ctx.define_result.get("largest_box", None) final_boxes, final_labels, final_scores, final_kws, final_raws, final_mods, dd_log = update_fallback_detection( ctx, pi_rgba, fallback_box, RTDETR_PROCESSOR, RTDETR_MODEL, DEVICE, RTDETR_CONF, final_boxes, final_labels, final_scores, final_kws, final_raws, final_mods, dd_log ) detection_result = { "status": "ok", "boxes": final_boxes, "labels": final_labels, "scores": final_scores, "final_keywords": final_kws, "raw_labels": final_raws, "models": final_mods } BOTTOM_CLOTHING_TYPES = ["jeans", "shorts", "skirt"] is_bottom_clothing = ctx.product_type.lower() in BOTTOM_CLOTHING_TYPES detection_log_schema = { "yolos_fashionpedia_model": { "person_list": {}, "product_type_list": {}, "head_list": {}, "shoes_list": {}, "clothing_features_list": {}, "artifacts_list": {} }, "rtdetr_model": { "person_list": {}, "product_type_list": {}, "head_list": {}, "shoes_list": {}, "clothing_features_list": {}, "artifacts_list": {} } } if not is_bottom_clothing: detection_log_schema["head_model"] = { "head_list": {} } for i in range(len(final_boxes)): keyword = final_kws[i] raw_label = final_raws[i] score = final_scores[i] model = final_mods[i] if is_bottom_clothing and model == "head_model": continue if model not in detection_log_schema: continue model_dict = detection_log_schema[model] if keyword in PERSON_LIST and "person_list" in model_dict: if keyword not in model_dict["person_list"]: model_dict["person_list"][keyword] = {} model_dict["person_list"][keyword][raw_label] = score elif keyword in PRODUCT_TYPE_LIST and "product_type_list" in model_dict: if keyword not in model_dict["product_type_list"]: model_dict["product_type_list"][keyword] = {} model_dict["product_type_list"][keyword][raw_label] = score elif keyword in HEAD_LIST and "head_list" in model_dict: if keyword not in model_dict["head_list"]: model_dict["head_list"][keyword] = {} model_dict["head_list"][keyword][raw_label] = score elif keyword in SHOES_LIST and "shoes_list" in model_dict: if keyword not in model_dict["shoes_list"]: model_dict["shoes_list"][keyword] = {} model_dict["shoes_list"][keyword][raw_label] = score elif keyword in CLOTHING_FEATURES_LIST and "clothing_features_list" in model_dict: if keyword not in model_dict["clothing_features_list"]: model_dict["clothing_features_list"][keyword] = {} model_dict["clothing_features_list"][keyword][raw_label] = score elif keyword in ARTIFACTS_LIST and "artifacts_list" in model_dict: if keyword not in model_dict["artifacts_list"]: model_dict["artifacts_list"][keyword] = {} model_dict["artifacts_list"][keyword][raw_label] = score for model in list(detection_log_schema.keys()): model_dict = detection_log_schema[model] for cat in list(model_dict.keys()): if not model_dict[cat]: del model_dict[cat] if not model_dict: del detection_log_schema[model] log_item["status"] = "ok" log_item["data"]["detection_result_log"] = detection_log_schema ctx.detection_result = detection_result batch_logs.append(log_item) processing_time = time.perf_counter() - start_time logging.log(LOG_LEVEL_MAP["SUCCESS"], f"{EMOJI_MAP['SUCCESS']} Completed {function_name} for {len(contexts)} items in {processing_time:.3f}s") return batch_logs def choose_color_for_feature_batch(contexts, batch_logs): function_name = "choose_color_for_feature_batch" start_time = time.perf_counter() logging.log(LOG_LEVEL_MAP["INFO"], f"{EMOJI_MAP['INFO']} Starting {function_name} for {len(contexts)} items") processed_count = 0 skipped_count = 0 error_count = 0 for ctx in contexts: it_ = { "image_url": ctx.url, "function": function_name } if ctx.skip_run or ctx.skip_processing: it_["status"] = "skipped" batch_logs.append(it_) skipped_count += 1 continue dr = ctx.detection_result if dr.get("status") != "ok": it_["status"] = "no_detection" batch_logs.append(it_) error_count += 1 continue bxs = dr["boxes"] kws = dr["final_keywords"] pt = ctx.product_type colors = [] for i, (bx1, by1, bx2, by2) in enumerate(bxs): kw = kws[i] if kw == pt: colors.append(BLUE_BOX_PRODUCT_TYPE) elif kw in HEAD_LIST: colors.append(GREEN_BOX_HEAD) elif kw in SHOES_LIST: colors.append(VIOLET_BOX_SHOES) elif kw in CLOTHING_FEATURES_LIST: colors.append(ORANGE_BOX_CLOTHING_FEATURES) elif kw in ARTIFACTS_LIST: colors.append(RED_BOX_ARTIFACTS) else: colors.append(BLUE_BOX_PRODUCT_TYPE) ctx.box_colors = colors it_["status"] = "ok" it_["data"] = { "color_assignments": { "total_boxes": len(bxs), "color_counts": { "product_type": colors.count(BLUE_BOX_PRODUCT_TYPE), "head": colors.count(GREEN_BOX_HEAD), "shoes": colors.count(VIOLET_BOX_SHOES), "clothing_features": colors.count(ORANGE_BOX_CLOTHING_FEATURES), "artifacts": colors.count(RED_BOX_ARTIFACTS) } } } batch_logs.append(it_) processed_count += 1 processing_time = time.perf_counter() - start_time logging.log(LOG_LEVEL_MAP["SUCCESS"], f"{EMOJI_MAP['SUCCESS']} Completed {function_name}: processed={processed_count}, skipped={skipped_count}, errors={error_count} in {processing_time:.3f}s") return batch_logs def adjust_blue_box_batch(contexts, batch_logs): function_name = "adjust_blue_box_batch" start_time = time.perf_counter() logging.log(LOG_LEVEL_MAP["INFO"], f"{EMOJI_MAP['INFO']} Starting {function_name} for {len(contexts)} items") processed_count = 0 skipped_count = 0 error_count = 0 for ctx in contexts: it_ = { "image_url": ctx.url, "function": function_name } if ctx.skip_run or ctx.skip_processing: it_["status"] = "skipped" batch_logs.append(it_) skipped_count += 1 continue dr = ctx.detection_result if dr.get("status") != "ok": it_["status"] = "no_detection" batch_logs.append(it_) error_count += 1 continue if "original" not in ctx.pil_img: it_["status"] = "error" it_["exception"] = "No original image found" batch_logs.append(it_) error_count += 1 continue pi_rgba, _, _ = ctx.pil_img["original"] largest_box = ctx.define_result.get("largest_box") if not largest_box: it_["status"] = "error" it_["exception"] = "No main box detected" batch_logs.append(it_) error_count += 1 continue final_boxes = dr["boxes"] final_kws = dr["final_keywords"] final_cols = ctx.box_colors bx1, by1, bx2, by2 = map(int, largest_box) largest_green = None largest_violet = None largest_blue = None areaG = 0 areaV = 0 areaB = 0 for i, (xx1, yy1, xx2, yy2) in enumerate(final_boxes): area_ = (xx2 - xx1) * (yy2 - yy1) col_ = final_cols[i] kw_ = final_kws[i] if kw_ in HEAD_LIST and area_ > areaG: areaG = area_ largest_green = (xx1, yy1, xx2, yy2) elif kw_ in SHOES_LIST and area_ > areaV: areaV = area_ largest_violet = (xx1, yy1, xx2, yy2) if col_ == BLUE_BOX_PRODUCT_TYPE and area_ > areaB: areaB = area_ largest_blue = (xx1, yy1, xx2, yy2) adjusted_box = largest_blue lines_update = {} adjustment_log = { "original_blue_box": largest_blue, "head_box": largest_green, "shoes_box": largest_violet, "adjustments": {} } if adjusted_box and areaB > 0: x1b, y1b, x2b, y2b = adjusted_box m = min(pi_rgba.width, pi_rgba.height) horiz_off = int(m * LEFT_RIGHT_BLUE_OFFSET) up_ofs = int(m * LOWER_UPPER_BLUE_OFFSET) dn_ofs = int(m * LOWER_UPPER_BLUE_OFFSET) x1_ext = max(x1b - horiz_off, bx1) x2_ext = min(x2b + horiz_off, bx2) y1_ext = y1b y2_ext = y2b adjustment_log["adjustments"]["horizontal_offset"] = horiz_off adjustment_log["adjustments"]["vertical_offset"] = {"up": up_ofs, "down": dn_ofs} if largest_green: gx1, gy1, gx2, gy2 = largest_green top_off = int(m * CREATE_UPPER_BLUE_LOWER_GREEN_RATIO) new_top = gy2 - top_off y1_ext = max(by1, new_top) adjustment_log["adjustments"]["head_top_offset"] = top_off else: y1_ext = max(y1_ext - up_ofs, by1) adjustment_log["adjustments"]["default_top_offset"] = up_ofs if largest_violet: vx1, vy1, vx2, vy2 = largest_violet bot_off = int(m * CREATE_LOWER_BLUE_LOWER_VIOLET_RATIO) new_bot = vy2 + bot_off y2_ext = min(by2, max(y2_ext, new_bot)) adjustment_log["adjustments"]["shoes_bottom_offset"] = bot_off else: y2_ext = min(y2_ext + dn_ofs, by2) adjustment_log["adjustments"]["default_bottom_offset"] = dn_ofs x1_ext = max(bx1, min(x1_ext, bx2)) x2_ext = max(bx1, min(x2_ext, bx2)) y1_ext = max(by1, min(y1_ext, by2)) y2_ext = max(by1, min(y2_ext, by2)) adjusted_box = (x1_ext, y1_ext, x2_ext, y2_ext) lines_update = { "left": f"{x1b}->{x1_ext}", "right": f"{x2b}->{x2_ext}", "upper": f"{y1b}->{y1_ext}", "lower": f"{y2b}->{y2_ext}" } adjustment_log["adjustments"]["before"] = [x1b, y1b, x2b, y2b] adjustment_log["adjustments"]["after"] = [x1_ext, y1_ext, x2_ext, y2_ext] largest_area = (bx2 - bx1) * (by2 - by1) adjusted_area = (x2_ext - x1_ext) * (y2_ext - y1_ext) adjustment_log["areas"] = { "largest_box_area": largest_area, "adjusted_area": adjusted_area } if largest_area > 0: ratio = (adjusted_area / float(largest_area)) * 100 adjustment_log["areas"]["ratio_percent"] = round(ratio, 2) if ratio < BLUE_BOX_FALLBACK_THRESHOLD: adjusted_box = (bx1, by1, bx2, by2) lines_update["fallback"] = ( f"Adjusted area {round(ratio,1)}% < {BLUE_BOX_FALLBACK_THRESHOLD}%, " "fallback to entire largest box." ) adjustment_log["fallback_used"] = True adjustment_log["fallback_reason"] = f"Area ratio {round(ratio,2)}% below threshold {BLUE_BOX_FALLBACK_THRESHOLD}%" else: adjustment_log["fallback_used"] = False ctx.adjusted_blue_box = adjusted_box it_["status"] = "ok" it_["lines_update"] = lines_update it_["data"] = {"adjustment_log": adjustment_log} batch_logs.append(it_) processed_count += 1 processing_time = time.perf_counter() - start_time logging.log(LOG_LEVEL_MAP["SUCCESS"], f"{EMOJI_MAP['SUCCESS']} Completed {function_name}: processed={processed_count}, skipped={skipped_count}, errors={error_count} in {processing_time:.3f}s") return batch_logs def draw_batch(contexts, batch_logs): function_name = "draw_batch" start_time = time.perf_counter() logging.log(LOG_LEVEL_MAP["INFO"], f"{EMOJI_MAP['INFO']} Starting {function_name} for {len(contexts)} items") processed_count = 0 skipped_count = 0 error_count = 0 for ctx in contexts: it_ = { "image_url": ctx.url, "function": function_name } if ctx.skip_run or ctx.skip_processing: it_["status"] = "skipped" batch_logs.append(it_) skipped_count += 1 continue dr = ctx.detection_result if dr.get("status") != "ok": it_["status"] = "no_detection" batch_logs.append(it_) error_count += 1 continue if "original" not in ctx.pil_img: it_["status"] = "error" it_["exception"] = "No original image found" batch_logs.append(it_) error_count += 1 continue pi_rgba, orig_filename, _ = ctx.pil_img["original"] final_boxes = dr["boxes"] final_kws = dr["final_keywords"] final_cols = ctx.box_colors largest_box = ctx.define_result.get("largest_box") box_drawn = 0 color_log = { "BLACK_BOX_PERSON": {"count": 0, "boxes": []}, "BLUE_BOX_PRODUCT_TYPE": {"count": 0, "boxes": []}, "GREEN_BOX_HEAD": {"count": 0, "boxes": []}, "VIOLET_BOX_SHOES": {"count": 0, "boxes": []}, "ORANGE_BOX_CLOTHING_FEATURES": {"count": 0, "boxes": []}, "RED_BOX_ARTIFACTS": {"count": 0, "boxes": []}, } apply_draw = Apply_Draw if apply_draw: d_ = ImageDraw.Draw(pi_rgba, mode="RGBA") for i, (bx1, by1, bx2, by2) in enumerate(final_boxes): c_ = final_cols[i] if c_ != BLUE_BOX_PRODUCT_TYPE: s_ = f"({bx1},{by1},{bx2},{by2})" d_.rectangle([bx1, by1, bx2, by2], outline=c_, width=2) box_drawn += 1 if c_ == GREEN_BOX_HEAD: color_log["GREEN_BOX_HEAD"]["count"] += 1 color_log["GREEN_BOX_HEAD"]["boxes"].append(s_) elif c_ == VIOLET_BOX_SHOES: color_log["VIOLET_BOX_SHOES"]["count"] += 1 color_log["VIOLET_BOX_SHOES"]["boxes"].append(s_) elif c_ == ORANGE_BOX_CLOTHING_FEATURES: color_log["ORANGE_BOX_CLOTHING_FEATURES"]["count"] += 1 color_log["ORANGE_BOX_CLOTHING_FEATURES"]["boxes"].append(s_) elif c_ == RED_BOX_ARTIFACTS: color_log["RED_BOX_ARTIFACTS"]["count"] += 1 color_log["RED_BOX_ARTIFACTS"]["boxes"].append(s_) if ctx.adjusted_blue_box: abx1, aby1, abx2, aby2 = ctx.adjusted_blue_box s_ = f"({abx1},{aby1},{abx2},{aby2})" d_.rectangle([abx1, aby1, abx2, aby2], outline=BLUE_BOX_PRODUCT_TYPE, width=2) color_log["BLUE_BOX_PRODUCT_TYPE"]["count"] += 1 color_log["BLUE_BOX_PRODUCT_TYPE"]["boxes"].append(s_) box_drawn += 1 if largest_box: lx1, ly1, lx2, ly2 = largest_box s_ = f"({lx1},{ly1},{lx2},{ly2})" d_.rectangle([lx1, ly1, lx2, ly2], outline=BLACK_BOX_PERSON, width=4) color_log["BLACK_BOX_PERSON"]["count"] += 1 color_log["BLACK_BOX_PERSON"]["boxes"].append(s_) box_drawn += 1 ctx.pil_img["original"] = [pi_rgba, orig_filename, None] it_["status"] = "ok" if apply_draw else "no_drawing" it_["boxes_drawn"] = box_drawn it_["data"] = { "colors": color_log, "draw_enabled": apply_draw, "boxes_count": { "total": len(final_boxes), "drawn": box_drawn } } batch_logs.append(it_) processed_count += 1 processing_time = time.perf_counter() - start_time logging.log(LOG_LEVEL_MAP["SUCCESS"], f"{EMOJI_MAP['SUCCESS']} Completed {function_name}: processed={processed_count}, skipped={skipped_count}, errors={error_count} in {processing_time:.3f}s") return batch_logs # ---------------------------------------------------------------------- # MAIN PIPELINE FUNCTION # ---------------------------------------------------------------------- def ensure_models_loaded(): import app app.ensure_models_loaded() pipeline_step = create_pipeline_step(ensure_models_loaded) @pipeline_step def bounding_box(contexts: List[ProcessingContext], batch_logs: List[dict] = None): if batch_logs is None: batch_logs = [] start_time = time.perf_counter() logging.log(LOG_LEVEL_MAP["INFO"], f"{EMOJI_MAP['INFO']} Starting bounding_box pipeline for {len(contexts)} items") from src.models import model_loader RTDETR_PROCESSOR = model_loader.RTDETR_PROCESSOR RTDETR_MODEL = model_loader.RTDETR_MODEL RTDETR_FULL_PRECISION = model_loader.RTDETR_FULL_PRECISION HEAD_PROCESSOR = model_loader.HEAD_PROCESSOR HEAD_MODEL = model_loader.HEAD_MODEL HEAD_DETECTION_FULL_PRECISION = model_loader.HEAD_DETECTION_FULL_PRECISION YOLOS_PROCESSOR = model_loader.YOLOS_PROCESSOR YOLOS_MODEL = model_loader.YOLOS_MODEL DEVICE = model_loader.DEVICE MODELS_LOADED = model_loader.MODELS_LOADED LOAD_ERROR = model_loader.LOAD_ERROR define_largest_box_batch( contexts, batch_logs, RTDETR_PROCESSOR, RTDETR_MODEL, RTDETR_FULL_PRECISION, DEVICE, MODELS_LOADED, LOAD_ERROR ) detect_batch( contexts, batch_logs, HEAD_PROCESSOR, HEAD_MODEL, HEAD_DETECTION_FULL_PRECISION, RTDETR_PROCESSOR, RTDETR_MODEL, RTDETR_FULL_PRECISION, YOLOS_PROCESSOR, YOLOS_MODEL, DEVICE, MODELS_LOADED, LOAD_ERROR ) choose_color_for_feature_batch(contexts, batch_logs) adjust_blue_box_batch(contexts, batch_logs) draw_batch(contexts, batch_logs) processing_time = time.perf_counter() - start_time logging.log(LOG_LEVEL_MAP["SUCCESS"], f"{EMOJI_MAP['SUCCESS']} Completed bounding_box pipeline for {len(contexts)} items in {processing_time:.3f}s") return batch_logs