nxp/YOLOv8

Overview

Ultralytics released Ultralytics YOLOv8 on January 10, 2023, delivering state‑of‑the‑art accuracy and speed. Building on the progress of earlier Ultralytics YOLO versions, it introduces improved features and optimizations that make it a strong choice for a wide range of object detection, segmentation and pose estimation tasks across many applications.

Key Features of Ultralytics YOLOv8

• Advanced Backbone and Neck Architectures: YOLOv8 incorporates modern backbone and neck designs that enhance feature extraction and overall detection performance.
• Anchor‑Free Split Ultralytics Head: By using an anchor‑free split head, YOLOv8 achieves higher accuracy and more efficient detection compared to traditional anchor‑based methods.
• Balanced Accuracy–Speed Optimization: YOLOv8 is engineered to deliver strong accuracy while maintaining real‑time speed, making it well‑suited for a wide range of real‑time detection applications.
• Multiple Pretrained Model Options: A variety of pretrained models are available, allowing users to choose the one that best matches their task requirements and performance needs.

Model Description

This is a repository that contains a set of quantized and compiled versions of Ultralytics YOLOv8 models optimized for Ara240 DNPU.

• Base Model: Ultralytics/YOLOv8
• Original Model Authors: Ultralytics
• Original License: AGPL-3.0
• Modified by: NXP

Modifications

This model is a derivative work with the following changes from the original:

• Quantization: INT8 calibrated using COCO val2017
• Compilation: Compiled for Ara240 DNPU
• Format: Converted to DVM format for NPU deployment

Original model available at: Ultralytics YOLOv8.

Performance Summary

Object Detection

Model	size (pixels)	FP32 mAPval 50-95	INT8 mAPval 50-95	Latency Ara240 (ms)	Performance Ara240 (inferences/s)	params (M)
YOLOv8n	640	37.3	35.5	3.15	316.99	3.2
YOLOv8s	640	44.9	43.4	6.72	148.72	11.2
YOLOv8m	640	50.2	48.7	17.60	56.80	25.9
YOLOv8l	640	52.9	51.4	35.96	27.80	43.7
YOLOv8x	640	53.9	49.7	56.02	17.84	68.2

• mAP^val values are for single-model single-scale on COCO val2017 dataset.

Segmentation

Model	size (pixels)	FP32 mAPmask 50-95	INT8 mAPmask 50-95	Latency Ara240 (ms)	Performance Ara240 (inferences/s)	params (M)
YOLOv8n-seg	640	30.5	29.18	4.13	241.88	3.4
YOLOv8s-seg	640	36.8	36.10	10.14	98.61	11.8
YOLOv8m-seg	640	40.8	39.71	25.07	39.88	27.3
YOLOv8l-seg	640	42.6	39.15	49.87	20.04	46
YOLOv8x-seg	640	43.4	39.6	78.91	12.67	71.8

• mAP^mask values are for single-model single-scale on COCO val2017 dataset.

Pose

Model	size (pixels)	FP32 mAPpose 50-95	INT8 mAPpose 50-95	Latency Ara240 (ms)	Performance Ara240 (inferences/s)	params (M)
YOLOv8n-pose	640	50.4	46.75	4.08	245.01	3.3
YOLOv8s-pose	640	60	53.66	8.07	123.88	11.6
YOLOv8m-pose	640	65	58.84	19.02	52.56	26.4
YOLOv8l-pose	640	67.6	62.84	36.85	27.13	44.4
YOLOv8x-pose	640	69.2	64.76	59.48	16.8	69.4

• mAP^pose values are for single-model single-scale on COCO val2017 dataset.

License

Ultralytics offers two licensing options to accommodate diverse use cases:

• AGPL-3.0 License: This OSI-approved open-source license is ideal for students and enthusiasts, promoting open collaboration and knowledge sharing. See the LICENSE file for more details.
• Enterprise License: Designed for commercial use, this license permits seamless integration of Ultralytics software and AI models into commercial goods and services, bypassing the open-source requirements of AGPL-3.0. If your scenario involves embedding our solutions into a commercial offering, reach out through Ultralytics Licensing.