nxp/YOLOX-Detection-Ara240

Overview

Megvii released YOLOvX in July 2021. The release represented an inflection point in the family of YOLO models, with major architectural changes such as Decoupled Head as well as Anchor-Free detection. The average precision (AP) achieved by these models surpassed the previous generations of YOLOv3–YOLOv5 models, while maintaining real-time speed for a variety of applications.

Key Features of YOLOX

• Decoupled Head: YOLOX lessens the tradeoff between classification and regression tasks by introducing a decoupled head. This decoupled head consists of 1x1 and 3x3 convolutions arranged in parallel. With decoupled head, model is able to accurately localize the objects while maintaining high class confidence.
• Anchor-free detection: In contrast to its predecessors YOLOv3–YOLOv5, an anchor-free architecture is proposed in YOLOX, in which only 4 parameters per location are predicted (two offsets, width, and height).
• Advanced Label Assignment: YOLOX introduces SimOTA as the label assignment strategy, which helps reduce training time and increase accuracy.
• Data Augmentation: To improve generalization error, YOLOX leverages MixUp and Mosaic strategies for data augmentation, thus bypassing dependency on ImageNet pretraining.

YOLOX model comparison on V100: accuracy (mAP), latency and number of parameters across different model sizes. Source: Megvii-BaseDetection/YOLOX

Model Description

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

• Base Model: Megvii/YOLOX
• Original Model Authors: Megvii Technology
• Original License: Apache-2.0
• 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 Ara240 NPU deployment

Original model available at: Megvii

Performance and Accuracy Summary

Object Detection

Model	size (pixels)	FP32 mAPval 50-95	INT8 mAPval 50-95	Latency Ara240 (ms)	Performance Ara240 (inferences/s)	params (M)
YOLOX-Tiny	416	32.8	30.57	2.48	402.98	5.06
YOLOX-s	640	40.5	38.34	6.52	153.25	9.0
YOLOX-m	640	46.9	44.13	15.89	62.89	25.3
YOLOX-l	640	49.7	47.04	31.51	31.72	54.2
YOLOX-x	640	51.1	48.43	59.69	16.75	99.1

• mAP^val values are obtained on COCO val2017 dataset.

License

This model is released under the Apache-2.0 License, the same license as the original Megvii-BaseDetection/YOLOX model.

Citation

• Z. Ge, S. Liu, F. Wang, Z. Li, and J. Sun, “YOLOX: Exceeding YOLO Series in 2021,” CoRR, vol. abs/2107.08430, 2021, [Online]. Available: https://arxiv.org/abs/2107.08430

If you use this model, please cite both this work and the original model:

@article{DBLP:journals/corr/abs-2107-08430,
  author       = {Zheng Ge and
                  Songtao Liu and
                  Feng Wang and
                  Zeming Li and
                  Jian Sun},
  title        = {{YOLOX:} Exceeding {YOLO} Series in 2021},
  journal      = {CoRR},
  volume       = {abs/2107.08430},
  year         = {2021},
  url          = {https://arxiv.org/abs/2107.08430},
  eprinttype   = {arXiv},
  eprint       = {2107.08430},
  timestamp    = {Tue, 05 Apr 2022 14:09:44 +0200},
  biburl       = {https://dblp.org/rec/journals/corr/abs-2107-08430.bib},
  bibsource    = {dblp computer science bibliography, https://dblp.org}
}