A PyTorch implementation of PDNet: Toward Better One-Stage Object Detection With Prediction Decoupling. IEEE Transactions on Image Processing, 2022.

We propose a Prediction-target-Decoupled Network (PDNet) for object detection, which decouples the prediction of different targets (i.e., object category and boundaries) into their respective proper positions. The figure below is a visualization of the decoupled regression predictions for localizing each object box boundary. This allows us to collect accurate predictions for these targets separately. For more details on Prediction Decoupling, please refer to our paper.

This implementation is based on maskrcnn-benchmark, so the installation procedure is pretty much the same. Please check INSTALL.md for details. Additionally, please remember to build the CUDA implementation of the prediction collection module with the following commands.

cd pdnet_core/pred_collect_ops
python setup.py build develop

We provide the following trained models with different backbones or training settings.

[1] 2x means the model is trained for 180k iterations ( $\approx$ 24 epochs) with multi-scale training ([640, 800]), while 2xx denotes using a wider training scale range [480, 960].
[2] With efficient prediction decoupling, our model has a similar inference time to FCOS while achieving significantly higher accuracy.

Evaluate the trained model on the minival split of MS COCO dataset:

python tools/test_net.py --config-file configs/pdnet/pdnet_R_50_FPN_1x.yaml --ckpt PDNet_R_50_FPN_1x.pth

• You can change the arguments of --config-file and --ckpt to evaluate different models.

We train our detection model on 4 GPUs with 16 images per minibatch using the following command:

python -m torch.distributed.launch --nproc_per_node=4 tools/train_net.py --config-file configs/pdnet/pdnet_R_50_FPN_1x.yaml

• We also provide configuration files with different backbones or training settings in configs/pdnet/.

Please cite our paper if this project helps your research.

@article{yang2022pdnet,
  title={PDNet: Toward Better One-Stage Object Detection With Prediction Decoupling},
  author={Yang, Li and Xu, Yan and Wang, Shaoru and Yuan, Chunfeng and Zhang, Ziqi and Li, Bing and Hu, Weiming},
  journal={IEEE Transactions on Image Processing},
  volume={31},
  pages={5121--5133},
  year={2022},
  publisher={IEEE}
}