TOOD: Task-aligned One-stage Object Detection

TOOD is an object detection model. We reproduced the model of the paper.

Notes:

• TOOD is trained on COCO train2017 dataset and evaluated on val2017 results of mAP(IoU=0.5:0.95).
• TOOD uses 8GPU to train 12 epochs.

GPU multi-card training

export CUDA_VISIBLE_DEVICES=0,1,2,3,4,5,6,7
python -m paddle.distributed.launch --gpus 0,1,2,3,4,5,6,7 tools/train.py -c configs/tood/tood_r50_fpn_1x_coco.yml --fleet

@inproceedings{feng2021tood,
    title={TOOD: Task-aligned One-stage Object Detection},
    author={Feng, Chengjian and Zhong, Yujie and Gao, Yu and Scott, Matthew R and Huang, Weilin},
    booktitle={ICCV},
    year={2021}
}