Awesome

OneNet: What Makes for End-to-End Object Detection?

Comparisons of different label assignment methods. H and W are height and width of feature map, respectively, K is number of object categories. Previous works on one-stage object detection assign labels by only position cost, such as (a) box IoU or (b) point distance between sample and ground-truth. In our method, however, (c) classification cost is additionally introduced. We discover that classification cost is the key to the success of end-to-end. Without classification cost, only location cost leads to redundant boxes of high confidence scores in inference, making NMS post-processing a necessary component.

Introduction

arxiv: OneNet: Towards End-to-End One-Stage Object Detection

paper: What Makes for End-to-End Object Detection?

Updates

• (28/06/2021) OneNet.RetinaNet and OneNet.FCOS on CrowdHuman are available.
• (27/06/2021) OneNet.RetinaNet and OneNet.FCOS are available.
• (11/12/2020) Higher Performance for OneNet is reported by disable gradient clip.

Comming

• Provide models and logs
• Support to caffe, onnx, tensorRT
• Support to MobileNet

Models on COCO

We provide two models

• dcn is for high accuracy
• nodcn is for easy deployment.

If download link is invalid, models and logs are also available in Github Release and Baidu Drive by code nhr8.

Notes

• We observe about 0.3 AP noise.
• The training time and inference time are on 8 NVIDIA V100 GPUs. We observe the same type of GPUs in different clusters may cost different time.
• We use the models pre-trained on imagenet using torchvision. And we provide torchvision's ResNet-18.pkl model. More details can be found in the conversion script.

Models on CrowdHuman

If download link is invalid, models and logs are also available in Github Release and Baidu Drive by code nhr8.

Notes

• The evalution code is built on top of cvpods.
• The default evaluation code in training should be ignored, since it only considers at most 100 objects in one image, while crowdhuman image contains more than 100 objects.
• The training time and inference time are on 8 NVIDIA V100 GPUs. We observe the same type of GPUs in different clusters may cost different time.
• More training steps are in the crowdhumantools.

Installation

The codebases are built on top of Detectron2 and DETR.

Requirements

• Linux or macOS with Python ≥ 3.6
• PyTorch ≥ 1.5 and torchvision that matches the PyTorch installation. You can install them together at pytorch.org to make sure of this
• OpenCV is optional and needed by demo and visualization

Steps

1. Install and build libs

git clone https://github.com/PeizeSun/OneNet.git
cd OneNet
python setup.py build develop

2. Link coco dataset path to OneNet/datasets/coco

mkdir -p datasets/coco
ln -s /path_to_coco_dataset/annotations datasets/coco/annotations
ln -s /path_to_coco_dataset/train2017 datasets/coco/train2017
ln -s /path_to_coco_dataset/val2017 datasets/coco/val2017

3. Train OneNet

python projects/OneNet/train_net.py --num-gpus 8 \
    --config-file projects/OneNet/configs/onenet.res50.dcn.yaml

4. Evaluate OneNet

python projects/OneNet/train_net.py --num-gpus 8 \
    --config-file projects/OneNet/configs/onenet.res50.dcn.yaml \
    --eval-only MODEL.WEIGHTS path/to/model.pth

5. Visualize OneNet

python demo/demo.py\
    --config-file projects/OneNet/configs/onenet.res50.dcn.yaml \
    --input path/to/images --output path/to/save_images --confidence-threshold 0.4 \
    --opts MODEL.WEIGHTS path/to/model.pth

License

OneNet is released under MIT License.

Citing

If you use OneNet in your research or wish to refer to the baseline results published here, please use the following BibTeX entries:

@InProceedings{peize2020onenet,
  title = 	 {What Makes for End-to-End Object Detection?},
  author =       {Sun, Peize and Jiang, Yi and Xie, Enze and Shao, Wenqi and Yuan, Zehuan and Wang, Changhu and Luo, Ping},
  booktitle = 	 {Proceedings of the 38th International Conference on Machine Learning},
  pages = 	 {9934--9944},
  year = 	 {2021},
  volume = 	 {139},
  series = 	 {Proceedings of Machine Learning Research},
  publisher =    {PMLR},
}