This repo accompanies the following paper:

Olly Styles, Tanaya Guha, Victor Sanchez, “Multi-Camera Trajectory Forecasting with Trajectory Tensors”, tPAMI 2021

Preprint available: [ArXiv link]

All experiments are conducted on the [WNMF dataset], which is required to run the code. Please note that the results in the paper use the updated annotations (available after 2021.08.10). If you requested the dataset before this date, please send a new dataset request to get the updated annotations.

Clone repo:

git clone https://github.com/olly-styles/trajectory-tensors

Create a virtual environment:

python3 -m venv venv
source venv/bin/activate

Install requirements:

pip install -r requirements.txt

After downloading the dataset, preprocess the data using the following command. Please note: This operation requires 85GB of storage. It will take a few minutes to run, depending on your system:

python preprocessing/preprocess_all_data.py

After completing this operation, all tests should pass. You can test that the repo has been installed correctly and the dataset has been processed by running the tests:

python -m pytest testing

All experiments from the paper for the which, when, and where problem formulations are found in the “experiments “directory. Note that the CNN-GRU trajectory tensor model loads weights from the trajectory tensor autoencoder. The weights are created using “experiments/autoencoder/autoencoder.py “

After running the experiments and saving prediction, scripts to compute metrics are found in the “evaluation “directory.