Alpha-IoU: A Family of Power Intersection over Union Losses for Bounding Box Regression

YOLOv5 with alpha-IoU losses implemented in PyTorch.

Example results on the test set of PASCAL VOC 2007 using YOLOv5s trained by the vanilla IoU loss (top row) and the alpha-IoU loss with alpha=3 (bottom row). The alpha-IoU loss performs better than the vanilla IoU loss because it can localize objects more accurately (image 1 and 2), thus can detect more true positive objects (image 3 to 5) and fewer false positive objects (image 6 and 7).

Example results on the val set of MS COCO 2017 using YOLOv5s trained by the vanilla IoU loss (top row) and the alpha-IoU loss with alpha=3 (bottom row). The alpha-IoU loss performs better than the vanilla IoU loss because it can localize objects more accurately (image 1), thus can detect more true positive objects (image 2 to 5) and fewer false positive objects (image 4 to 7). Note that image 4 and 5 detect both more true positive and fewer false positive objects.

Citation

If you use our method, please consider citing:

@inproceedings{Jiabo_Alpha-IoU,
  author    = {He, Jiabo and Erfani, Sarah and Ma, Xingjun and Bailey, James and Chi, Ying and Hua, Xian-Sheng},
  title     = {Alpha-IoU: A Family of Power Intersection over Union Losses for Bounding Box Regression},
  booktitle = {NeurIPS},
  year      = {2021},
}

Modifications

This repository is a fork of ultralytics/yolov5, with an implementation of alpha-IoU losses while keeping the code as close to the original as possible.

Alpha-IoU Losses

Alpha-IoU losses can be configured in Line 131 of utils/loss.py, functionesd as 'bbox_alpha_iou'. The alpha values and types of losses (e.g., IoU, GIoU, DIoU, CIoU) can be selected in this function, which are defined in utils/general.py. Note that we should use a small constant epsilon to avoid torch.pow(0, alpha) or denominator=0.

Install

Python>=3.6.0 is required with all requirements.txt installed including PyTorch>=1.7:

$ git clone https://github.com/Jacobi93/Alpha-IoU
$ cd Alpha-IoU
$ pip install -r requirements.txt

Configurations

Configuration files can be found in data. We do not change either 'voc.yaml' or 'coco.yaml' used in the original repository. However, we could do more experiments. E.g.,

voc25.yaml # randomly use 25% PASCAL VOC as the training set
voc50.yaml # randomly use 50% PASCAL VOC as the training set

Code for generating different small training sets is in generate_small_sets.py. Code for generating different noisy labels is in generate_noisy_labels.py, and we should change the 'img2label_paths' function in utils/datasets.py accordingly.

Implementation Commands

For detailed installation instruction and network training options, please take a look at the README file or issue of ultralytics/yolov5. Following are sample commands we used for training and testing YOLOv5 with alpha-IoU, with more samples in instruction.txt.

python train.py --data voc.yaml --hyp hyp.scratch.yaml --cfg yolov5s.yaml --batch-size 64 --epochs 300 --device '0'
python test.py --data voc.yaml --img 640 --conf 0.001 --weights 'runs/train/voc_yolov5s_iou/weights/best.pt' --device '0'
python detect.py --source ../VOC/images/detect500 --weights 'runs/train/voc_yolov5s_iou/weights/best.pt' --conf 0.25

We can also randomly generate some images for detection and visualization results in generate_detect_images.py.

Pretrained Weights

Here are some pretrained models using the configurations in this repository, with alpha=3 in all experiments. Details of these pretrained models can be found in runs/train. All results are tested using 'weights/best.pt' for each experiment. It is a very simple yet effective method so that people is able to quickly apply our method to existing models following the 'bbox_alpha_iou' function in utils/general.py. Note that YOLOv5 has been updated for many versions and all pretrained models in this repository are obtained based on the YOLOv5 version 4.0, where details of all versions for YOLOv5 can be found. Researchers are also welcome to apply our method to other object detection models, e.g., Faster R-CNN, DETR, etc.

• The IoU loss for YOLOv5s on PASCAL VOC
• The alpha-IoU loss for YOLOv5s on PASCAL VOC
• The DIoU loss for YOLOv5s on PASCAL VOC
• The alpha-DIoU loss for YOLOv5s on PASCAL VOC