Home

Awesome

VTE

[Paper, ICCV 2021]

Poster

poster

Setup

Getting Started

conda install pytorch torchvision torchaudio cudatoolkit=10.2 -c pytorch

https://github.com/kregmi/VTE.git
cd VTE

Code

GeoTemporalFeatureLearning:

Code to conduct geo-temporal feature learning. The encoder consists of code for 2D CNN backbone. The temporalAttention consists of the implementation of transformer-based attention module.

Training is done in two stages.

First, train the `encoder` module using the `main.py` file.
Save the encoder featues using the `test_bdd.py` file.
Second, train the `temporalAttention` module using the `train.py` file.
Use `eval.py` to evaluate the trained model. 

Additional instruction is provided within each module's README.

TrajectorySmoothingNetwork:

Code to train/test the trajectory smoothing network.

`train.py` contains the code for training. 
`eval.py` contains the code for testing.

Dataset

BDD videos are downloaded from the official BDD website.

The corresponding Google StreetView Images are downloaded using Google Cloud Platform.

Please contact us for additional instruction to obtain the dataset used in this work.

Citation

If you find our works useful for your research, please cite our work:

Questions

Please contact: krishna.regmi7@gmail.com