Spatial-Temporal Transformer for Dynamic Scene Graph Generation

Pytorch Implementation of our paper Spatial-Temporal Transformer for Dynamic Scene Graph Generation accepted by ICCV2021. We propose a Transformer-based model STTran to generate dynamic scene graphs of the given video. STTran can detect the visual relationships in each frame.

The introduction video is available now: https://youtu.be/gKpnRU8btLg

About the code We run the code on a single RTX2080ti for both training and testing. We borrowed some code from Yang's repository and Zellers' repository.

Usage

We use python=3.6, pytorch=1.1 and torchvision=0.3 in our code. First, clone the repository:

git clone https://github.com/yrcong/STTran.git

We borrow some compiled code for bbox operations.

cd lib/draw_rectangles
python setup.py build_ext --inplace
cd ..
cd fpn/box_intersections_cpu
python setup.py build_ext --inplace

For the object detector part, please follow the compilation from https://github.com/jwyang/faster-rcnn.pytorch We provide a pretrained FasterRCNN model for Action Genome. Please download here and put it in

fasterRCNN/models/faster_rcnn_ag.pth

Dataset

We use the dataset Action Genome to train/evaluate our method. Please process the downloaded dataset with the Toolkit. The directories of the dataset should look like:

|-- action_genome
    |-- annotations   #gt annotations
    |-- frames        #sampled frames
    |-- videos        #original videos

In the experiments for SGCLS/SGDET, we only keep bounding boxes with short edges larger than 16 pixels. Please download the file object_bbox_and_relationship_filtersmall.pkl and put it in the dataloader

Train

You can train the STTran with train.py. We trained the model on a RTX 2080ti:

• For PredCLS:

python train.py -mode predcls -datasize large -data_path $DATAPATH 

• For SGCLS:

python train.py -mode sgcls -datasize large -data_path $DATAPATH 

• For SGDET:

python train.py -mode sgdet -datasize large -data_path $DATAPATH 

Evaluation

You can evaluate the STTran with test.py.

• For PredCLS (trained Model):

python test.py -m predcls -datasize large -data_path $DATAPATH -model_path $MODELPATH

• For SGCLS (trained Model): :

python test.py -m sgcls -datasize large -data_path $DATAPATH -model_path $MODELPATH

• For SGDET (trained Model): :

python test.py -m sgdet -datasize large -data_path $DATAPATH -model_path $MODELPATH

Citation

If our work is helpful for your research, please cite our publication:

@inproceedings{cong2021spatial,
  title={Spatial-Temporal Transformer for Dynamic Scene Graph Generation},
  author={Cong, Yuren and Liao, Wentong and Ackermann, Hanno and Rosenhahn, Bodo and Yang, Michael Ying},
  booktitle = {International Conference on Computer Vision (ICCV)},
  year={2021}
  url={https://arxiv.org/abs/2107.12309}
}

Help

When you have any question/idea about the code/paper. Please comment in Github or send us Email. We will reply as soon as possible.