Lightweight human detection models generated on high-quality human data sets. It can detect objects with high accuracy and speed in a total of 13 classes: Body, BodyWithWheelchair, BodyWithCrutches,Head, Face, Eye, Nose, Mouth, Ear, Hand, Hand-Left, Hand-Right, Foot(Feet). Even the classification problem is being attempted to be solved by object detection. There is no need to perform any complex affine transformations or other processing for pre-processing and post-processing of input images. In addition, the resistance to Motion Blur, Gaussian noise, contrast noise, backlighting, and halation is quite strong because it was trained only on images with added photometric noise for all images in the MS-COCO subset of the image set. In addition, about half of the image set was annotated by me with the aspect ratio of the original image substantially destroyed. I manually annotated all images in the dataset by myself. The model is intended to use real-world video for inference and has enhanced resistance to all kinds of noise. Probably stronger than any known model. However, the quality of the known data set and my data set are so different that an accurate comparison of accuracy is not possible.
Don't be ruled by the curse of mAP.
The use of CD-COCO: Complex Distorted COCO database for Scene-Context-Aware computer vision has also greatly improved resistance to various types of noise.
- Global distortions
- Noise
- Contrast
- Compression
- Photorealistic Rain
- Photorealistic Haze
- Motion-Blur
- Defocus-Blur
- Backlight illumination
- Local distortions
- Motion-Blur
- Defocus-Blur
- Backlight illumination
- COCO-Hand http://vision.cs.stonybrook.edu/~supreeth/COCO-Hand.zip
- CD-COCO: Complex Distorted COCO database for Scene-Context-Aware computer vision
- I am adding my own enhancement data to COCO-Hand and re-annotating all images. In other words, only COCO images were cited and no annotation data were cited.
- I have no plans to publish my own dataset.
Halfway compromises are never acceptable.
| Class Name | Class ID |
|---|---|
| Body | 0 |
| BodyWithWheelchair | 1 |
| BodyWithCrutches | 2 |
| Head | 3 |
| Face | 4 |
| Eye | 5 |
| Nose | 6 |
| Mouth | 7 |
| Ear | 8 |
| Hand | 9 |
| Hand-Left | 10 |
| Hand-Right | 11 |
| Foot (Feet) | 12 |
-
Python 3.10
-
onnx 1.14.1+
-
onnxruntime-gpu v1.16.1 (TensorRT Execution Provider Enabled Binary. See: onnxruntime-gpu v1.16.1 + CUDA 11.8 + TensorRT 8.5.3 build (RTX3070))
-
opencv-contrib-python 4.9.0.80+
-
numpy 1.24.3
-
TensorRT 8.5.3-1+cuda11.8
# Common ############################################ pip install opencv-contrib-python numpy onnx # For ONNX ########################################## pip uninstall onnxruntime onnxruntime-gpu pip install onnxruntime or pip install onnxruntime-gpu
-
Demonstration of models with built-in post-processing (Float32/Float16)
usage: demo_yolov9_onnx_wholebody13.py \ [-h] \ [-m MODEL] \ (-v VIDEO | -i IMAGES_DIR) \ [-ep {cpu,cuda,tensorrt}] \ [-it] \ [-dvw] \ [-dwk] \ [-dlr] options: -h, --help show this help message and exit -m MODEL, --model MODEL ONNX/TFLite file path for YOLOv9. -v VIDEO, --video VIDEO Video file path or camera index. -i IMAGES_DIR, --images_dir IMAGES_DIR jpg, png images folder path. -ep {cpu,cuda,tensorrt}, \ --execution_provider {cpu,cuda,tensorrt} Execution provider for ONNXRuntime. -it {fp16,int8}, --inference_type {fp16,int8} Inference type. Default: fp16 -dvw, --disable_video_writer Disable video writer. Eliminates the file I/O load associated with automatic recording to MP4. Devices that use a MicroSD card or similar for main storage can speed up overall processing. -dwk, --disable_waitKey Disable cv2.waitKey(). When you want to process a batch of still images, disable key-input wait and process them continuously. -dlr, --disable_left_and_right_hand_discrimination_mode Disable left and right hand discrimination mode. -
YOLOv9-Wholebody13 - N - Swish/SiLU (PINTO original implementation, 2.4 MB)
Class Images Instances P R mAP50 mAP50-95 all 2385 69591 0.678 0.474 0.528 0.333 body 2385 12199 0.714 0.630 0.691 0.469 body_with_wheelchair 2385 182 0.696 0.824 0.845 0.673 body_with_crutches 2385 103 0.597 0.893 0.892 0.743 head 2385 10343 0.805 0.718 0.778 0.526 face 2385 5561 0.846 0.638 0.699 0.446 eye 2385 5211 0.567 0.198 0.216 0.079 nose 2385 4818 0.666 0.298 0.339 0.158 mouth 2385 3936 0.602 0.282 0.302 0.118 ear 2385 4874 0.683 0.303 0.347 0.169 hand 2385 7791 0.777 0.406 0.531 0.286 hand_left 2385 3938 0.638 0.274 0.394 0.216 hand_right 2385 3853 0.611 0.268 0.367 0.207 foot 2385 6782 0.613 0.433 0.468 0.240 -
YOLOv9-Wholebody13 - N - ReLU (PINTO original implementation, 2.4 MB, For INT8/QAT)
Class Images Instances P R mAP50 mAP50-95 all 2385 69591 0.675 0.47 0.523 0.329 body 2385 12199 0.687 0.627 0.674 0.449 body_with_wheelchair 2385 182 0.671 0.868 0.870 0.682 body_with_crutches 2385 103 0.559 0.883 0.861 0.747 head 2385 10343 0.787 0.702 0.758 0.511 face 2385 5561 0.825 0.637 0.696 0.442 eye 2385 5211 0.613 0.201 0.232 0.083 nose 2385 4818 0.681 0.306 0.343 0.160 mouth 2385 3936 0.637 0.28 0.310 0.118 ear 2385 4874 0.671 0.302 0.346 0.166 hand 2385 7791 0.773 0.388 0.516 0.276 hand_left 2385 3938 0.626 0.263 0.383 0.209 hand_right 2385 3853 0.635 0.249 0.361 0.200 foot 2385 6782 0.615 0.403 0.446 0.227 -
YOLOv9-Wholebody13 - T - Swish/SiLU
Class Images Instances P R mAP50 mAP50-95 all 2385 69591 0.772 0.583 0.652 0.442 body 2385 12199 0.796 0.749 0.803 0.603 body_with_wheelchair 2385 182 0.835 0.923 0.950 0.845 body_with_crutches 2385 103 0.694 0.913 0.936 0.867 head 2385 10343 0.855 0.794 0.850 0.609 face 2385 5561 0.861 0.755 0.812 0.559 eye 2385 5211 0.682 0.270 0.317 0.125 nose 2385 4818 0.759 0.430 0.489 0.244 mouth 2385 3936 0.705 0.373 0.412 0.179 ear 2385 4874 0.753 0.399 0.454 0.239 hand 2385 7791 0.888 0.545 0.697 0.420 hand_left 2385 3938 0.736 0.446 0.579 0.355 hand_right 2385 3853 0.748 0.419 0.551 0.344 foot 2385 6782 0.730 0.558 0.622 0.354 -
YOLOv9-Wholebody13 - S - Swish/SiLU
Class Images Instances P R mAP50 mAP50-95 all 2385 69591 0.838 0.653 0.726 0.514 body 2385 12199 0.835 0.805 0.856 0.683 body_with_wheelchair 2385 182 0.883 0.951 0.980 0.889 body_with_crutches 2385 103 0.912 0.932 0.956 0.922 head 2385 10343 0.879 0.846 0.891 0.660 face 2385 5561 0.894 0.784 0.837 0.625 eye 2385 5211 0.750 0.327 0.406 0.169 nose 2385 4818 0.820 0.491 0.567 0.310 mouth 2385 3936 0.765 0.424 0.494 0.237 ear 2385 4874 0.790 0.476 0.539 0.300 hand 2385 7791 0.913 0.657 0.795 0.516 hand_left 2385 3938 0.828 0.577 0.706 0.466 hand_right 2385 3853 0.833 0.554 0.689 0.455 foot 2385 6782 0.789 0.662 0.729 0.446 -
YOLOv9-Wholebody13 - M - Swish/SiLU
Class Images Instances P R mAP50 mAP50-95 all 2385 69591 0.848 0.674 0.753 0.540 body 2385 12199 0.835 0.820 0.865 0.697 body_with_wheelchair 2385 182 0.900 0.934 0.971 0.888 body_with_crutches 2385 103 0.890 0.947 0.975 0.928 head 2385 10343 0.874 0.863 0.903 0.683 face 2385 5561 0.896 0.836 0.878 0.670 eye 2385 5211 0.769 0.342 0.435 0.186 nose 2385 4818 0.849 0.532 0.616 0.351 mouth 2385 3936 0.805 0.459 0.540 0.269 ear 2385 4874 0.808 0.494 0.570 0.322 hand 2385 7791 0.909 0.667 0.817 0.550 hand_left 2385 3938 0.861 0.590 0.742 0.510 hand_right 2385 3853 0.853 0.583 0.730 0.498 foot 2385 6782 0.779 0.696 0.751 0.471 -
YOLOv9-Wholebody13 - C - Swish/SiLU
Class Images Instances P R mAP50 mAP50-95 all 2385 69591 0.861 0.676 0.761 0.553 body 2385 12199 0.850 0.832 0.874 0.716 body_with_wheelchair 2385 182 0.898 0.934 0.968 0.892 body_with_crutches 2385 103 0.967 0.942 0.974 0.937 head 2385 10343 0.880 0.870 0.909 0.694 face 2385 5561 0.900 0.833 0.875 0.676 eye 2385 5211 0.778 0.347 0.452 0.197 nose 2385 4818 0.856 0.537 0.624 0.363 mouth 2385 3936 0.810 0.468 0.553 0.277 ear 2385 4874 0.802 0.505 0.576 0.331 hand 2385 7791 0.909 0.646 0.820 0.569 hand_left 2385 3938 0.869 0.581 0.749 0.524 hand_right 2385 3853 0.870 0.570 0.736 0.516 foot 2385 6782 0.805 0.720 0.777 0.499 -
YOLOv9-Wholebody13 - E - Swish/SiLU
Class Images Instances P R mAP50 mAP50-95 all 2385 69591 0.886 0.723 0.806 0.607 body 2385 12199 0.890 0.859 0.904 0.773 body_with_wheelchair 2385 182 0.924 0.951 0.974 0.911 body_with_crutches 2385 103 0.971 0.969 0.991 0.962 head 2385 10343 0.898 0.888 0.926 0.735 face 2385 5561 0.897 0.874 0.907 0.727 eye 2385 5211 0.807 0.438 0.540 0.251 nose 2385 4818 0.894 0.607 0.690 0.423 mouth 2385 3936 0.835 0.551 0.632 0.345 ear 2385 4874 0.839 0.584 0.662 0.403 hand 2385 7791 0.930 0.663 0.853 0.627 hand_left 2385 3938 0.899 0.622 0.789 0.587 hand_right 2385 3853 0.882 0.61 0.780 0.576 foot 2385 6782 0.850 0.778 0.837 0.574 -
Pre-Process
To ensure fair benchmark comparisons with YOLOX,
BGR to RGB conversion processingandnormalization by division by 255.0are added to the model input section. In addition, aresizing processfor input images has been added to improve operational flexibility. Thus, in any model, inferences can be made at any image size. The string1x3x{H}x{W}at the end of the file name does not indicate the input size of the image, but the processing resolution inside the model. Therefore, the smaller the values of{H}and{W}, the lower the computational cost and the faster the inference speed. Models with larger values of{H}and{W}increase the computational cost and decrease the inference speed. Since the concept is different from the resolution of an image, any size image can be batch processed. e.g. 240x320, 480x640, 720x1280, ...
-
Post-Process
Because I add my own post-processing to the end of the model, which can be inferred by TensorRT, CUDA, and CPU, the benchmarked inference speed is the end-to-end processing speed including all pre-processing and post-processing. EfficientNMS in TensorRT is very slow and should be offloaded to the CPU.
-
NMS default parameter
param value note max_output_boxes_per_class 20 Maximum number of outputs per class of one type. 20indicates that the maximum number of people detected is20, the maximum number of heads detected is20, and the maximum number of hands detected is20. The larger the number, the more people can be detected, but the inference speed slows down slightly due to the larger overhead of NMS processing by the CPU. In addition, as the number of elements in the final output tensor increases, the amount of information transferred between hardware increases, resulting in higher transfer costs on the hardware circuit. Therefore, it would be desirable to set the numerical size to the minimum necessary.iou_threshold 0.40 A value indicating the percentage of occlusion allowed for multiple bounding boxes of the same class. 0.40is excluded from the detection results if, for example, two bounding boxes overlap in more than 41% of the area. The smaller the value, the more occlusion is tolerated, but over-detection may increase.score_threshold 0.25 Bounding box confidence threshold. Specify in the range of 0.00to1.00. The larger the value, the stricter the filtering and the lower the NMS processing load, but in exchange, all but bounding boxes with high confidence values are excluded from detection. This is a parameter that has a very large percentage impact on NMS overhead. -
Change NMS parameters
Use PINTO0309/sam4onnx to rewrite the
NonMaxSuppressionparameter in the ONNX file.For example,
pip install onnxsim==0.4.33 \ && pip install -U simple-onnx-processing-tools \ && pip install -U onnx \ && python -m pip install -U onnx_graphsurgeon \ --index-url https://pypi.ngc.nvidia.com ### max_output_boxes_per_class ### Example of changing the maximum number of detections per class to 100. sam4onnx \ --op_name main01_nonmaxsuppression13 \ --input_onnx_file_path yolov9_e_wholebody13_post_0245_1x3x544x960.onnx \ --output_onnx_file_path yolov9_e_wholebody13_post_0245_1x3x544x960.onnx \ --input_constants main01_max_output_boxes_per_class int64 [100] ### iou_threshold ### Example of changing the allowable area of occlusion to 20%. sam4onnx \ --op_name main01_nonmaxsuppression13 \ --input_onnx_file_path yolov9_e_wholebody13_post_0245_1x3x544x960.onnx \ --output_onnx_file_path yolov9_e_wholebody13_post_0245_1x3x544x960.onnx \ --input_constants main01_iou_threshold float32 [0.20] ### score_threshold ### Example of changing the bounding box score threshold to 15%. sam4onnx \ --op_name main01_nonmaxsuppression13 \ --input_onnx_file_path yolov9_e_wholebody13_post_0245_1x3x544x960.onnx \ --output_onnx_file_path yolov9_e_wholebody13_post_0245_1x3x544x960.onnx \ --input_constants main01_score_threshold float32 [0.15]
-
Post-processing structure
PyTorch alone cannot generate this post-processing. For operational flexibility,
EfficientNMSis not used.
-
-
INT8 quantization (YOLOv9-QAT)
If this work has contributed in any way to your research or business, I would be happy to be cited in your literature.
@software{YOLOv9-Wholebody13,
author={Katsuya Hyodo},
title={Lightweight human detection models generated on high-quality human data sets. It can detect objects with high accuracy and speed in a total of 13 classes: Body, BodyWithWheelchair, BodyWithCrutches, Head, Face, Eye, Nose, Mouth, Ear, Hand, Hand-Left, Hand-Right, Foot.},
url={https://github.com/PINTO0309/PINTO_model_zoo/tree/main/454_YOLOv9-Wholebody13},
year={2024},
month={7},
doi={10.5281/zenodo.10229410}
}I am very grateful for their excellent work.
-
COCO-Hand
https://vision.cs.stonybrook.edu/~supreeth/
@article{Hand-CNN, title={Contextual Attention for Hand Detection in the Wild}, author={Supreeth Narasimhaswamy and Zhengwei Wei and Yang Wang and Justin Zhang and Minh Hoai}, booktitle={International Conference on Computer Vision (ICCV)}, year={2019}, url={https://arxiv.org/pdf/1904.04882.pdf} }
-
CD-COCO: Complex Distorted COCO database for Scene-Context-Aware computer vision
@INPROCEEDINGS{10323035, author={Beghdadi, Ayman and Beghdadi, Azeddine and Mallem, Malik and Beji, Lotfi and Cheikh, Faouzi Alaya}, booktitle={2023 11th European Workshop on Visual Information Processing (EUVIP)}, title={CD-COCO: A Versatile Complex Distorted COCO Database for Scene-Context-Aware Computer Vision}, year={2023}, volume={}, number={}, pages={1-6}, doi={10.1109/EUVIP58404.2023.10323035} }
-
YOLOv9
https://github.com/WongKinYiu/yolov9
@article{wang2024yolov9, title={{YOLOv9}: Learning What You Want to Learn Using Programmable Gradient Information}, author={Wang, Chien-Yao and Liao, Hong-Yuan Mark}, booktitle={arXiv preprint arXiv:2402.13616}, year={2024} }
-
YOLOv9-QAT