Language:
🇺🇸
🇨🇳
«YOLOv2» reproduced the paper "YOLO9000: Better, Faster, Stronger"
- Train using the
VOC07+12 trainvaldataset and test using theVOC2007 Testdataset with an input size of640x640. Given the result as follows
| Original (darknet) | tztztztztz/yolov2.pytorch | zjykzj/YOLOv2(This) | zjykzj/YOLOv2(This) | |
|---|---|---|---|---|
| ARCH | YOLOv2 | YOLOv2 | YOLOv2 | YOLOv2-Fast |
| GFLOPs | / | / | 69.5 | 48.5 |
| DATASET(TRAIN) | VOC TRAINVAL 2007+2012 | VOC TRAINVAL 2007+2012 | VOC TRAINVAL 2007+2012 | VOC TRAINVAL 2007+2012 |
| DATASET(VAL) | VOC TEST 2007 | VOC TEST 2007 | VOC TEST 2007 | VOC TEST 2007 |
| INPUT_SIZE | 416x416 | 416x416 | 640x640 | 640x640 |
| PRETRAINED | TRUE | TRUE | FALSE | FALSE |
| VOC AP[IoU=0.50:0.95] | / | / | 44.3 | 29.8 |
| VOC AP[IoU=0.50] | 76.8 | 72.7 | 75.1 | 62.6 |
- Train using the
COCO train2017dataset and test using theCOCO val2017dataset with an input size of640x640. Given the result as follows (Note: The results of the original paper were evaluated on theCOCO test-dev2015dataset)
| Original (darknet) | zjykzj/YOLOv2(This) | zjykzj/YOLOv2(This) | |
|---|---|---|---|
| ARCH | YOLOv2 | YOLOv2 | YOLOv2-Fast |
| GFLOPs | / | 69.7 | 48.8 |
| DATASET(TRAIN) | / | COCO TRAIN2017 | COCO TRAIN2017 |
| DATASET(VAL) | / | COCO VAL2017 | COCO VAL2017 |
| INPUT_SIZE | 416x416 | 640x640 | 640x640 |
| PRETRAINED | TRUE | FALSE | FALSE |
| COCO AP[IoU=0.50:0.95] | 21.6 | 28.6 | 20.1 |
| COCO AP[IoU=0.50] | 44.0 | 50.7 | 41.2 |
- [2024/05/04]v1.0.0. Refactoring YOLOv2 project, integrating yolov5 v7.0, reimplementing YOLOv2/YOLOv2-fast and YOLOv2Loss.
- [2023/07/16]v0.3.0. Add ultralytics/yolov5(485da42) transforms.
- [2023/06/28]v0.2.1. Refactor data module.
- [2023/05/21]v0.2.0. Reconstructed loss function and add Darknet53 as a backbone.
- [2023/05/09]v0.1.2. Add COCO dataset result and update VOC dataset training results.
- [2023/05/03]v0.1.1. Fix target transform and update
yolov2_voc.cfgandyolov2-tiny_voc.cfgtraining results for VOC2007 Test. - [2023/05/02]v0.1.0. Complete YOLOv2 training/evaluation/prediction, while providing the evaluation results of VOC2007 Test.
YOLOv2 has made more innovations on the basis of YOLOv1. For the network, it has created Darknet-19; For the loss function, it adds anchor box settings to help network training with more fine-grained features. Compared with YOLOv1, YOLOv2 is more modern and high-performance.
This repository references many repositories implementations, including tztztztztz/yolov2.pytorch and yjh0410/yolov2-yolov3_PyTorch, as well as zjykzj/YOLOv3.
Note: the latest implementation of YOLOv2 in our warehouse is entirely based on ultralytics/yolov5 v7.0
pip3 install -r requirements.txtOr use docker container
docker run -it --runtime nvidia --gpus=all --shm-size=16g -v /etc/localtime:/etc/localtime -v $(pwd):/workdir --workdir=/workdir --name yolov2 ultralytics/yolov5:v7.0python3 train.py --data VOC.yaml --weights "" --cfg yolov2_voc.yaml --img 640 --device 0 --yolov2loss
python3 train.py --data VOC.yaml --weights "" --cfg yolov2-fast_voc.yaml --img 640 --device 0 --yolov2loss
python3 train.py --data coco.yaml --weights "" --cfg yolov2.yaml --img 640 --device 0 --yolov2loss
python3 train.py --data coco.yaml --weights "" --cfg yolov2-fast.yaml --img 640 --device 0 --yolov2loss# python3 val.py --weights runs/train/voc/exp/weights/best.pt --data VOC.yaml --img 640 --device 0
yolov2_voc summary: 53 layers, 50645053 parameters, 0 gradients, 69.5 GFLOPs
Class Images Instances P R mAP50 mAP50-95: 100%|██████████| 155/155 00:41
all 4952 12032 0.735 0.711 0.751 0.443
Speed: 0.1ms pre-process, 3.1ms inference, 1.3ms NMS per image at shape (32, 3, 640, 640)
# python3 val.py --weights runs/train/voc/exp3/weights/best.pt --data VOC.yaml --img 640 --device 0
yolov2-fast_voc summary: 33 layers, 42367485 parameters, 0 gradients, 48.5 GFLOPs
Class Images Instances P R mAP50 mAP50-95: 100%|██████████| 155/155 00:37
all 4952 12032 0.626 0.612 0.626 0.298
Speed: 0.1ms pre-process, 2.3ms inference, 1.5ms NMS per image at shape (32, 3, 640, 640)
# python3 val.py --weights runs/train/coco/exp/weights/best.pt --data coco.yaml --img 640 --device 0
yolov2 summary: 53 layers, 50952553 parameters, 0 gradients, 69.7 GFLOPs
Class Images Instances P R mAP50 mAP50-95: 100%|██████████| 157/157 00:57
all 5000 36335 0.627 0.48 0.507 0.286
Speed: 0.1ms pre-process, 3.1ms inference, 2.0ms NMS per image at shape (32, 3, 640, 640)
# python3 val.py --weights runs/train/coco/exp2/weights/best.pt --data coco.yaml --img 640 --device 0
yolov2-fast summary: 33 layers, 42674985 parameters, 0 gradients, 48.8 GFLOPs
Class Images Instances P R mAP50 mAP50-95: 100%|██████████| 157/157 00:53
all 5000 36335 0.549 0.402 0.412 0.201
Speed: 0.1ms pre-process, 2.4ms inference, 2.1ms NMS per image at shape (32, 3, 640, 640)python3 detect.py --weights runs/yolov2_voc.pt --source ./assets/voc2007-test/
python3 detect.py --weights runs/yolov2_coco.pt --source ./assets/coco/
- zhujian - Initial work - zjykzj
- zjykzj/vocdev
- zjykzj/YOLOv3
- zjykzj/anchor-boxes
- ultralytics/yolov5
- AlexeyAB/darknet
- tztztztztz/yolov2.pytorch
- yjh0410/yolov2-yolov3_PyTorch
Anyone's participation is welcome! Open an issue or submit PRs.
Small note:
- Git submission specifications should be complied with Conventional Commits
- If versioned, please conform to the Semantic Versioning 2.0.0 specification
- If editing the README, please conform to the standard-readme specification.
Apache License 2.0 © 2023 zjykzj