YOLOXY: Common 2D computer vision task with Anchor Free YOLO.

Roadmap

• Object Detection
• Keypoint Detection
• Multi Objects Tracking(ByteTrack)
• Model Prune

Object Detection	Object Detection With Keypoints

Other Demo

Face Detection With Keypoints	Face Detection With Keypoints

Benchmark

Nano Size

Model	size	mAPval 0.5:0.95	Params(M)	FLOPs(G)	Speed(ms) b32 fp32 RTX2080Ti
YOLOv5n(v6.2)	640	28.0	1.9	4.5	0.8
YOLOv6-N	640	36.3	4.3	11.1	1.38
YOLOX-Nano	416	25.8	0.91	1.08	-
YOLOXY-N	640	33.9	1.97	4.3	0.8

Small Size

Model	size	mAPval 0.5:0.95	Params(M)	FLOPs(G)	Speed(ms) b32 fp32 RTX2080Ti
YOLOv5s(v6.2)	640	37.4	7.23	16.53	1.4
YOLOv6-S	640	43.8	17.2	44.2	3.49
YOLOX-s	640	40.5	9.0	26.8	-
PP-YOLOE+_s	640	43.7	7.93	17.36	-
RTMDet-s	640	44.5	8.89	14.8	-
YOLOXY-S	640	42.0	7.71	16.7	1.4

Quick Start

Installation

Python>=3.7.0, PyTorch>=1.8.1

git clone https://github.com/jamjamjon/YOLOXY.git  # clone
cd YOLOXY
pip install -r requirements.txt  # install

Inference Detection

python tools/detect.py --weights yoloxy-n.pt --conf 0.5 --source 0  		# webcam
								 img.jpg  	# image
								 video.mp4  	# video
								 ./data/images/  # image directory
								 rtsp://admin:admin12345@192.168.0.177/h264  # RTSP stream

Inference Detection With Keypoints

--is-coco means to draw skeleton for COCO human keypoints(17). It is optional.

python tools/detect.py --weights yoloxy-s-coco-kpts.pt --conf 0.4 --kpts-conf 0.5 --is-coco --source 0  		# webcam
											   	     img.jpg  	# image
											             video.mp4  	# video
											  	     ./data/images/  # image directory
											  	     rtsp://admin:admin12345@192.168.0.177/h264  # RTSP stream

Inference With Multi Objects Tracking(MOT)

python tools/detect.py --weights yoloxy-n.pt --tracking --source rtsp://admin:admin12345@192.168.0.188/h264 

Test mAP

python tools/val.py --weights yoloxy-n.pt --data data/datasets/coco.yaml --img 640 --conf 0.001 --iou 0.65

Test Speed

python tools/val.py --weights yoloxy-n.pt --task speed --data data/datasets/coco.yaml --img 640 --batch 32

Export ONNX

python tools/export.py  --weights yoloxy-s.pt --img 640 --simplify  --include onnx

Training

Directories Structure

parent
├── YOLOXY
     └── data
    	  └── datasets
	  	└──your-dataset.yaml  ← step 1. prepare custom dataset.yaml 
└── datasets
     └── your-datasets  ←  step 2. custom dataset (images and labels)
   	  └── images
   	  └── labels

same as YOLOv5, check this

Train From Stratch

# Single GPU
python tools/train.py --data data/datasets/your-custom-dataset.yaml --cfg models/cfg/yoloxy-n.yaml --batch-size 32

# Multi-GPU
python -m torch.distributed.run -nproc_per_node 4 tools/train.py --data data/datasets/your-custom-dataset.yaml --cfg models/cfg/yoloxy-n.yaml --batch 128 --device 0,1,2,3

Transfer Learning

# Single GPU
python tools/train.py --data data/datasets/your-custom-dataset.yaml --weights yoloxy-n.pt --batch-size 32

# Multi-GPU
python -m torch.distributed.run -nproc_per_node 4 tools/train.py --data data/datasets/your-custom-dataset.yaml --weights yoloxy-n.pt --batch 128 --device 0,1,2,3

Training With Keypoints

Keypoints Label Format 1: kpt_x, kpt_y, kpt_visibility(conf)

# class_id x_center y_center width height kpt1_x kpt1_y kpt1_visibility kpt2_x kpt2_y kpt2_visibility ... kptn_x kptn_y kptn_visibility (normalized, 0-1)
0  0.03369140625 0.4786450662739323 0.0205078125 0.03829160530191458 0.0 0.0 0.0 0.04082421875 0.4736480117820324 1.0 0.03767578125 0.48089396170839466 2.0 0.033203125 0.48838880706921944 2.0 0.0403271484375 0.48813843888070696 2.0 	# one object

Keypoints Label Format 2: kpt_x, kpt_y

# class_id x_center y_center width height kpt1_x kpt1_y kpt2_x kpt2_y ... kptn_x kptn_y (normalized, 0-1)
0  0.03369140625 0.4786450662739323 0.0205078125 0.03829160530191458 0.0317119140625 0.4736480117820324 0.04082421875 0.4736480117820324 0.03767578125 0.48089396170839466 0.033203125 0.48838880706921944 0.0403271484375 0.48813843888070696   	# one object
1  0.09765625 0.47201767304860087 0.029296875 0.05154639175257732 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0	# another object

step 1. Prepare Dataset.yaml

parent
├── YOLOXY
     └── data
   	  └── datasets
	  	└──your-dataset.yaml  ← put here

And modify configs

# datasets path
path: ../datasets/coco-kpts  # dataset root dir
train: images/train2017  # train images 
val: images/val2017 # val images 

# Classes & Keypoints
nc: 1  # number of classes
names: ['person']
nk: 17   # number of keypoints
kpt_lr_flip_idx: [0, 2, 1, 4, 3, 6, 5, 8, 7, 10, 9, 12, 11, 14, 13, 16, 15]   # left-right flip for kpts, required when using ==> hyp['fliplr']

step 2. Prepare Dataset

parent
├── YOLOXY
     └── data
    	  └── datasets
	  	└──your-dataset.yaml  
└── datasets
      └── your-datasets  ←  put here
   	  └── images
   	  └── labels

step 3. Training

# Single GPU
python tools/train.py --data data/datasets/your-custom-dataset.yaml --weights yoloxy-s-coco-kpts.pt --batch 32

# Multi-GPU
python -m torch.distributed.run -nproc_per_node 4 tools/train.py --data data/datasets/your-custom-dataset.yaml --weights yoloxy-s-coco-kpts.pt --batch 128 --device 0,1,2,3

References

• https://github.com/ultralytics/yolov5
• https://github.com/Megvii-BaseDetection/YOLOX
• https://github.com/PaddlePaddle/PaddleDetection
• https://github.com/TexasInstruments/edgeai-yolov5/tree/yolo-pose