PointDETR

pytorch implementation of the paper, points as queries: weakly semi-supervised object detection by points

Under development .. comming soon! including:

* pretrained PointDETR at 20%  
* 20% bbox + 80% pseudo-bbox annotation file (PointDETR.json)

Requirements

This work is tested under:

ubuntu 18.04
python 3.6.9
torch 1.5.1
cuda 10.1

Installation

pip install -r requirements.txt

Instructions

0. Data Preparation

COCO dataset ./datasets/COCO
20% image ids in ./datasets/annoted_img_ids.py && ./cvpods/datasets/annoted_img_ids.py

1. Train PointDETR by 20% bbox

python3 -m torch.distributed.launch --nproc_per_node=8 --use_env main.py --coco_path ./datasets/COCO --partial_training_data --output_dir ./ckpt-ps/point-detr-9x --epochs 108 --lr_drop 72 --data_augment --position_embedding sine --warm_up --multi_step_lr

2. Generate 80% pseudo-bbox

python3 main.py --coco_path ./datasets/COCO --generate_pseudo_bbox --generated_anno PointDETR --position_embedding sine --resume ./ckpt-ps/point-detr-9x/baseline-checkpoint0107.pth

------- Student Model -------

Install cvpods

3. Train the student model with 20% bbox + 80% pseudo-bbox

cd ./cvpods/playground/detection/coco/fcos-20p-pointdetr
pods_train --num-gpus 8 --dir .

4. (optional) Train the student model with 20% bbox only.

cd ./cvpods/playground/detection/coco/fcos-20p-no_teacher
pods_train --num-gpus 8 --dir .

Citation

If this work helps your research / work, please consider citing:

@inproceedings{chen2021points,
  title={Points as queries: Weakly semi-supervised object detection by points},
  author={Chen, Liangyu and Yang, Tong and Zhang, Xiangyu and Zhang, Wei and Sun, Jian},
  booktitle={Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition},
  pages={8823--8832},
  year={2021}
}

Aknowledgement

This repo is built on the cvpods and DETR

cenchaojun/PointDETR