Forecasting Time-to-Collision from Monocular Video: Feasibility, Dataset and Challenges

This repository contains the official pytorch implementation for "Forecasting Time-to-Collision from Monocular Video: Feasibility, Dataset and Challenges". Please also check out our project page here. If you find this code useful, please cite our paper:

@article{Manglik2019, 
  archivePrefix = {arXiv}, 
  arxivId = {1903.09102}, 
  author = {Manglik, Aashi and Weng, Xinshuo and Ohn-bar, Eshed and Kitani, Kris}, 
  eprint = {1903.09102}, 
  journal = {arXiv:1903.09102}, 
  title = {{Forecasting Time-to-Collision from Monocular Video: Feasibility, Dataset and Challenges}}, 
  url = {https://arxiv.org/pdf/1903.09102.pdf}, 
  year = {2019} 
}

Dataset

Here is link to our dataset. The data is stored .mat format.

S.No.	Folder	Recordings
1	mats_nov	Left images, Right images, 3D point cloud, Calibration
2	mats_dec	Left images, Right images, 3D point cloud, Calibration
3	mat_stereo_camera	Left images, Right images, Depth Maps from Stereo Camera

Requirements

• Python 3.6
• PyTorch tested on 0.4.0

Installation

Setup virtualenv

Demo

1. Download the trained model:
   vgg_on_voc800
   6Image6s_027

2. Download preprocessed test data in .h5 format:
   6ImageTest.h5

Bounding Boxes

We used Faster RCNN to get the bounding boxes of people. Please see preprocessing/getBoundingBoxes.

Implementing in ROS

Please see test/multi_stream_prediction.py