Official Pytorch implementation for DRG: Dual Relation Graph for Human-Object Interaction Detection (ECCV 2020).

See the project page for more details. Please contact Jiarui Xu (jiaruixu@vt.edu) if you have any questions related to implementation details.

This codebase was tested with Python 3.6, Pytorch 1.0 from a nightly release, CUDA 10.0, and CentOS 7.4.1708.

Please check INSTALL.md for installation instructions.

Download V-COCO and HICO-DET data. Setup HICO-DET evaluation code.

bash ./scripts/download_dataset.sh 
bash ./scripts/download_data.sh

1. Download DRG detections and data

   bash ./scripts/download_drg_detection.sh

2. Evaluate on VCOCO

   python tools/vcoco_compute_mAP.py \
       --dataset_name vcoco_test \
       --detection_file output/VCOCO/detection_merged_human_object_app.pkl

3. Evaluate on HICO-DET

   cd Data/ho-rcnn
   matlab -r "Generate_detection('COCO'); quit"
   cd ../../

4. Evaluate on HICO-DET finetuned detection

   cd Data/ho-rcnn
   matlab -r "Generate_detection('finetune'); quit"
   cd ../../

0. Down pre-trained Faster R-CNN model weights for initialization

   bash ./scripts/download_frcnn.sh

1. Train on V-COCO

   bash ./scripts/train_VCOCO.sh

2. Train on HICO-DET

   bash ./scripts/train_HICO.sh

1. Test on V-COCO

   bash ./scripts/test_VCOCO.sh $APP_ITER_NUMBER $HUMAN_SP_ITER_NUMBER $OBJECT_SP_ITER_NUMBER

2. Test on HICO-DET

   bash ./scripts/test_HICO.sh $APP_ITER_NUMBER $HUMAN_SP_ITER_NUMBER $OBJECT_SP_ITER_NUMBER

3. Test on HICO-DET w/ a fine-tined detector

   bash ./scripts/test_HICO_ft.sh

NOTE: If you wish the use the same detector for a fair comparison, see here.

Download DRG trained weights.

bash ./scripts/download_drg_models.sh

For a simple demo, you can try

python demo/demo_obj_det.py

Currently, we only support Faster R-CNN with ResNet-R50-FPN backbone.

If you find this code useful for your research, please consider citing the following papers:

@inproceedings{Gao-ECCV-DRG,
    author    = {Gao, Chen and Xu, Jiarui and Zou, Yuliang and Huang, Jia-Bin}, 
    title     = {DRG: Dual Relation Graph for Human-Object Interaction Detection}, 
    booktitle = {European Conference on Computer Vision},
    year      = {2020}
}

@inproceedings{gao2018ican,
    author    = {Gao, Chen and Zou, Yuliang and Huang, Jia-Bin}, 
    title     = {iCAN: Instance-Centric Attention Network for Human-Object Interaction Detection}, 
    booktitle = {British Machine Vision Conference},
    year      = {2018}
}

This code follows the implementation architecture of maskrcnn-benchmark, iCAN and No Frills.