cyc-cp

This repository provides the benchmark for Cycling Close Pass Near Miss (Cyc-CP).

Installation

1. Clone the repo: git clone https://github.com/SustainableMobility/cyc-cp.git
2. Install the package:
   1. cd cyc-cp
   2. pip install -e .

How To Use

1. cd cyc-cp
2. Train crnn or i3d model following:
   1. python ./cnm/scene_lvl/i3d/i3d.py --csv_data_path path/to/csv_data --image_data_path path/to/image_data --exp_data_dir path/to/save/exp_data
   2. python ./cnm/scene_lvl/crnn/crnn.py --csv_data_path path/to/csv_data --image_data_path path/to/image_data --exp_data_dir path/to/save/exp_data
   • where the meaning of the arguments can be found in the code help. Specifically,
     • --csv_data_path: The file path of the .csv file with dataset info.
     • --image_data_path: The directory path saving all video frames.
     • --exp_data_dir: The directory to save results to.

Dataset Preparation

• Victorian On-bike Cycling (legacy): available on Monash Bridges
• Victorian On-bike Cycling (ongoing): under collection ...
• CARLA (simulation): available on Monash Bridges
• NuScences: available on https://www.nuscenes.org/

Hardware Requirements Summary

• Disk: to save all datasets about > 2TB disk space is required.
• RAM and GPU: (only tested on Victorian On-bike Cycling (legacy))
  • Scene-level:
    • I3D: (Batch_size:16, image_size: 256x342, frames: [-5, 15])
      • GPU memory: 7.5GB, CPU memory: 3 GB (RTX-2080 has 8 GB memory, so that’s why [-5, 15] frames are included in a video clip.)
    • CRNN: (Batch_size: 16, image_size: 224x224, frame: [20,25])
      • GPU memory: 3 GB, CPU memory: 6 GB
  • Instance-level (TODO)

More Notes about the project can be found in the shared google doc.