README.md

---
size_categories:
- 100K<n<1M
task_categories:
- video-classification
title: 'Trokens: Semantic-Aware Relational Trajectory Tokens Dataset'
tags:
- computer-vision
- action-recognition
- few-shot-learning
- video-understanding
- point-tracking
viewer: false
license: cc-by-nc-4.0
---

# Trokens: Semantic-Aware Relational Trajectory Tokens for Few-shot Action Recognition

This contains the preprocessed data for "Trokens: Semantic-Aware Relational Trajectory Tokens for Few-Shot Action Recognition" (ICCV 2025).

[**Paper**](https://arxiv.org/abs/2508.03695) | [**Project Page**](https://trokens-iccv25.github.io/) | [**Code**](https://github.com/pulkitkumar95/trokens)

## Dataset Overview

This dataset provides semantic-aware relational trajectory tokens (Trokens) extracted from multiple action recognition datasets, specifically for few-shot action recognition tasks. The dataset includes semantically meaningful point trajectories extracted using CoTracker3 and DINOv2 features, along with few-shot episode split information.

## Dataset Structure

The dataset contains two main components:

### 1. Point Tracking Data (`cotracker3_bip_fr_32/`)

Each dataset is present in a zip file. To unzip the dataset, run the following command:
```bash
cd cotracker3_bip_fr_32
unzip *.zip
```

Semantic point trajectories extracted using CoTracker3 with bipartite clustering on DINOv2 features:

```
cotracker3_bip_fr_32/
└── {dataset_name}/
    └── feat_dump/
        └── {video_name}.pkl
```

Each pickle file contains:
- **`pred_tracks`**: Tracked point coordinates across frames [T, N, 2]
- **`pred_visibility`**: Visibility mask for each point [T, N]  
- **`obj_ids`**: Object/cluster IDs for each point [N]
- **`point_queries`**: Original query point indices [N]
  
It also contains **`vid_info`**, which contains the video information of the video the points were extracted:
- **`fps`**: FPS at which the video was processed for point tracking.
- **`height`**: Height of the video.
- **`width`**: Width of the video.



### 2. Few-shot Split Information (`few_shot_info/`)

Data splits for few-shot learning evaluation across multiple datasets.

## Point Extraction Details

Code for extraction can be found on the GitHub repo [here](https://github.com/pulkitkumar95/trokens/tree/main/point_tracking). Some details are provided below.

### Semantic Point Tracking
- **Method**: CoTracker3 with semantic clustering on DINOv2 features
- **Clustering**: Bipartite clustering for semantic entity detection
- **Parameters**: 
  - Clustering method: `bipartite`
  - Number of frames for clustering: 32
  - Points filtered based on spatial proximity to remove redundancy

### Video Processing
- **Frame Rate**: 
  - Most datasets: 10 fps
  - Something Something V2 (SSV2): 12 fps (original video fps)
- **Point Filtering**: Redundant points removed based on spatial proximity
- **GPU Acceleration**: CUDA support for efficient processing

### Key Features
- Robust point tracking across video frames using CoTracker3
- Semantic point extraction through clustering on DINOv2 features
- Point filtering to remove redundant tracks
- Support for different clustering strategies and parameters

## Supported Datasets

The point tracking data is available for few shot splits of multiple action recognition datasets:
- **Something Something V2 (SSV2)**
- **Kinetics**
- **UCF-101**
- **HMDB-51**
- **Finegym**

## Usage

### Loading Point Tracking Data

```python
import pickle
import numpy as np

# Load point tracking data for a video
with open('cotracker3_bip_fr_32/{dataset}/{video_name}.pkl', 'rb') as f:
    data = pickle.load(f)
    
pred_tracks = data['pred_tracks']      # [T, N, 2] - point coordinates
pred_visibility = data['pred_visibility']  # [T, N] - visibility mask
obj_ids = data['obj_ids']              # [N] - cluster/object IDs
point_queries = data['point_queries']  # [N] - query point indices
```

### Loading Few-shot Splits

```python
# Load few-shot episode information
# (Structure depends on specific dataset format)
```

## Applications

This dataset is designed for:
- **Few-shot Action Recognition**: Training models with limited labeled examples
- **Video Understanding**: Learning from semantic-aware relational trajectory tokens (Trokens)
- **Point Tracking Research**: Semantic point trajectory analysis
- **Action Recognition**: General video classification tasks

## Technical Details

### Dependencies
- PyTorch
- NumPy  
- Pandas
- Einops
- CoTracker3 (for point tracking)
- DINOv2 (for feature extraction)

### Point Extraction Pipeline
1. **Feature Extraction**: DINOv2 features computed for video frames
2. **Semantic Clustering**: Bipartite clustering to identify semantic entities
3. **Point Sampling**: Points sampled from cluster centers
4. **Trajectory Tracking**: CoTracker3 used to track points across frames
5. **Post-processing**: Redundant points filtered based on spatial proximity

## Citation

If you use this dataset in your research, please cite our papers:

```bibtex
@inproceedings{kumar2025trokens,
  title={Trokens: Semantic-Aware Relational Trajectory Tokens for Few-Shot Action Recognition},
  author={Kumar, Pulkit and Huang, Shuaiyi and Walmer, Matthew and Rambhatla, Sai Saketh and Shrivastava, Abhinav},
  booktitle={International Conference on Computer Vision},
  year={2025}
}

@inproceedings{kumar2024trajectory,
  title={Trajectory-aligned Space-time Tokens for Few-shot Action Recognition},
  author={Kumar, Pulkit and Padmanabhan, Namitha and Luo, Luke and Rambhatla, Sai Saketh and Shrivastava, Abhinav},
  booktitle={European Conference on Computer Vision},
  pages={474--493},
  year={2024},
  organization={Springer}
}
```

## Authors

[**Pulkit Kumar***](https://www.cs.umd.edu/~pulkit/)<sup>1</sup> · [**Shuaiyi Huang***](https://shuaiyihuang.github.io/)<sup>1</sup> · [**Matthew Walmer**](https://www.cs.umd.edu/~mwalmer/)<sup>1</sup> · [**Sai Saketh Rambhatla**](https://rssaketh.github.io)<sup>1,2</sup> · [**Abhinav Shrivastava**](http://www.cs.umd.edu/~abhinav/)<sup>1</sup>

<sup>1</sup>University of Maryland, College Park&emsp;&emsp;&emsp;&emsp;<sup>2</sup>GenAI, Meta<br>
<sup>*Equal contribution</sup>

## License

This dataset is licensed under the [CC-BY-NC-4.0 License](https://creativecommons.org/licenses/by-nc/4.0/).

## Acknowledgments

This dataset is built upon:
- [CoTracker](https://github.com/facebookresearch/co-tracker): For robust point tracking
- [TATs](https://github.com/pulkitkumar95/tats): Trajectory-aligned Space-time Tokens for Few-shot Action Recognition
- [DINOv2](https://github.com/facebookresearch/dinov2): For semantic feature extraction


We thank the authors for making their code publicly available.