/YOLOv3

YOLOv3: An Incremental Improvement

Primary LanguagePythonApache License 2.0Apache-2.0

Language: 🇺🇸 🇨🇳

«YOLOv3» reproduced the paper "YOLOv3: An Incremental Improvement"

  • Train using the COCO train2017 dataset and test using the COCO val2017 dataset with an input size of 416x416. give the result as follows (No version of the COCO dataset used in the paper was found)
Original (darknet) DeNA/PyTorch_YOLOv3 zjykzj/YOLOv3(This)
ARCH YOLOv3 YOLOv3 YOLOv3
COCO AP[IoU=0.50:0.95] 0.310 0.311 0.400
COCO AP[IoU=0.50] 0.553 0.558 0.620

Table of Contents

Latest News

  • [2024/05/19]v5.1. Optimize YOLOv3Loss. Use BCELoss instead of MSELoss for confidence loss calculation.
  • [2024/05/09]v5.0. Refactoring YOLOv3 project, integrating yolov5 v7.0, reimplementing YOLOv3/YOLOv3-fast and YOLOv3Loss.
  • [2023/07/19]v4.0. Add ultralytics/yolov5(485da42) transforms and support AMP training.
  • [2023/06/22]v3.2. Remove Excess Code and Implementation.
  • [2023/06/22]v3.1. Reconstruct DATA Module and Preprocessing Module.
  • [2023/05/24]v3.0. Refer to zjykzj/YOLOv2 to reconstruct the entire project and train Pascal VOC and COCO datasets with YOLOv2Loss.
  • [2023/04/16]v2.0. Fixed preprocessing implementation, YOLOv3 network performance close to the original paper implementation.
  • [2023/02/16]v1.0. implementing preliminary YOLOv3 network training and inference implementation.

Background

The purpose of creating this warehouse is to better understand the YOLO series object detection network. Note: The realization of the project depends heavily on the implementation of DeNA/PyTorch_YOLOv3 and NVIDIA/apex

Note: the latest implementation of YOLOv3 in our warehouse is entirely based on ultralytics/yolov5 v7.0

Installation

pip3 install -r requirements.txt

Or use docker container

docker run -it --runtime nvidia --gpus=all --shm-size=16g -v /etc/localtime:/etc/localtime -v $(pwd):/workdir --workdir=/workdir --name yolov3 ultralytics/yolov5:v7.0

Usage

Train

python3 train.py --data VOC.yaml --weights "" --cfg yolov3_voc.yaml --img 640 --device 0 --yolov3loss
python3 train.py --data VOC.yaml --weights "" --cfg yolov3-fast_voc.yaml --img 640 --device 0 --yolov3loss
python3 train.py --data coco.yaml --weights "" --cfg yolov3_coco.yaml --img 640 --device 0 --yolov3loss
python3 train.py --data coco.yaml --weights "" --cfg yolov3-fast_coco.yaml --img 640 --device 0 --yolov3loss

Eval

# python3 val.py --weights runs/yolov3_voc.pt --data VOC.yaml --device 0
yolov3_voc summary: 198 layers, 67238145 parameters, 0 gradients, 151.5 GFLOPs
                 Class     Images  Instances          P          R      mAP50   mAP50-95: 100%|██████████| 155/155 01:04
                   all       4952      12032      0.811      0.742      0.816      0.568
             aeroplane       4952        285      0.938      0.791      0.897      0.607
Speed: 0.1ms pre-process, 6.8ms inference, 1.6ms NMS per image at shape (32, 3, 640, 640)
# python3 val.py --weights runs/yolov3-fast_voc.pt --data VOC.yaml --device 0
yolov3-fast_voc summary: 108 layers, 39945921 parameters, 0 gradients, 76.0 GFLOPs
                 Class     Images  Instances          P          R      mAP50   mAP50-95: 100%|██████████| 155/155 00:52
                   all       4952      12032      0.734      0.704      0.745       0.45
             aeroplane       4952        285      0.759      0.747      0.796      0.427
Speed: 0.1ms pre-process, 4.3ms inference, 1.6ms NMS per image at shape (32, 3, 640, 640)
# python3 val.py --weights runs/yolov3_coco.pt --data coco.yaml --device 0
yolov3_coco summary: 198 layers, 67561245 parameters, 0 gradients, 152.5 GFLOPs
Speed: 0.1ms pre-process, 6.9ms inference, 2.1ms NMS per image at shape (32, 3, 640, 640)
 Average Precision  (AP) @[ IoU=0.50:0.95 | area=   all | maxDets=100 ] = 0.400
 Average Precision  (AP) @[ IoU=0.50      | area=   all | maxDets=100 ] = 0.620
# python3 val.py --weights runs/yolov3-fast_coco.pt --data coco.yaml --device 0
yolov3-fast_coco summary: 108 layers, 40269021 parameters, 0 gradients, 77.0 GFLOPs
Speed: 0.1ms pre-process, 4.4ms inference, 2.2ms NMS per image at shape (32, 3, 640, 640)
 Average Precision  (AP) @[ IoU=0.50:0.95 | area=   all | maxDets=100 ] = 0.329
 Average Precision  (AP) @[ IoU=0.50      | area=   all | maxDets=100 ] = 0.560

Predict

python3 detect.py --weights runs/yolov3_voc.pt --source ./assets/voc2007-test/

python3 detect.py --weights runs/yolov3_coco.pt --source ./assets/coco/

Maintainers

  • zhujian - Initial work - zjykzj

Thanks

Contributing

Anyone's participation is welcome! Open an issue or submit PRs.

Small note:

License

Apache License 2.0 © 2022 zjykzj