Augmented Self-Mask Attention Transformer for Naturalistic Driving Action Recognition

The 1nd Place Submission to The 8th NVIDIA AI City Challenge (2024) Track 3: Naturalistic Driving Action Recognition.

Data Structure

Aicity2024-track3/data
|--extracted_features/      # the extracted feats for dataset A1&A2
|   |--A1/
|   |--A2/
|--raw_videos/              # original videos
|   |--A1/                  
|       |--user_id_xxx/
|           |--xxx.MP4
|   |--A2/                  
|   |--labels&instructions/ 
|       |--A1/   
|           |--xxx.csv          
|--crop_videos/             # videos of crop human, structure like raw_videos
|   |--A1/
|       |--user_id_xxx/
|           |--xxx.mp4
|   |--A2/
|--splited_videos/          # videos splited by label
|   |--A1/
|       |--xxx.MP4
            ...
|       |--splited_videos_label.csv  
|--label_A1-train_A1-val.json
|--label_submit.json

Workflow

The workflow for training action classification model is as follow:

1. Dataset Preparation
2. Feature Extraction
3. AMA
4. Post Process

Example for Test on Dataset B

1. Test Example

Contact

If you have any questions, feel free to contact Tiantian Zhang ( [email protected] / [email protected] ).