OAD benchmark

A small framework created to compare some SOTA online action detection methods, with focus on appliance in ADAS and ADS.

Setup instructions

Consider using conda or Docker to isolate your other projects or applications you use. Take a note, that Python must be 3.11 and Torch 2.9.0 for this setup.

1. Install the proper Torch wheel for your hardware:
   • CUDA / CPU
   • XPU
2. Run python setup/setup_dependencies.py to install MMAction2 and other dependencies from requirements.txt
3. (optional) Download ROAD_Waymo from here
   • Download videos and road_waymo_trainval_v1.0.json with gsutil, and place these files in ./data/road_waymo

     mkdir "data/road_waymo"
     cd "data/road_waymo"
     gcloud auth login
     gsutil -m cp -r "gs://waymo_open_dataset_road_plus_plus/road_waymo_trainval_v1.0.json" "gs://waymo_open_dataset_road_plus_plus/videos" .
     cd ../..

     You need to register yourself and accept terms of use before downloading.
4. Run python setup/get_road.py
   • ./data/ file structure should be following (including ROAD_Waymo):

     road-waymo/
         - road_trainval_v1.0.json
         - videos/
             - 2014-06-25-16-45-34_stereo_centre_02.mp4
             - 2014-06-26-09-53-12_stereo_centre_02.mp4
             - 2014-07-14-14-49-50_stereo_centre_01.mp4
             ...
     road-waymo/
         - road_waymo_trainval_v1.0.json
         - videos/
             - Train_00000.mp4
             - Train_00001.mp4
             - Train_00002.mp4
             ...
     

Including own elements

Feel free to include own methods or datasets to this framework. You can find more info in CONTRIBUTING.md guide.

Tested methods

This benchmark contains comparison of 4 LSTR family methods.

• TeSTrA
• CMeRT
• MAT
• MiniROAD