VINAY-UMRETHE/Dataset-SigMamba-V1-Large

---
tags:
- sigmamba
- video-anomaly-detection
- siglip2
- mamba
- features
task_categories:
- video-classification
license: cc-by-sa-4.0
---

# SigMamba-V1-Large-Features

<p align="center">
<img src="https://img.shields.io/badge/License-CC%20BY--SA%204.0-4c1?style=for-the-badge" alt="License">
<img src="https://img.shields.io/badge/Dataset-SigMamba--V1--Large-7952b3?style=for-the-badge" alt="Dataset">
</p>

This dataset contains visual features encoded from the dataset using the **SigLIP 2** vision encoder. These features are specifically prepared for **Weakly Supervised Video Anomaly Detection (WSVAD**) tasks, such as training the original **[VINAY-UMRETHE/SigMamba-V1-Large](https://huggingface.co/VINAY-UMRETHE/SigMamba-V1-Large)** model.

---

## Dataset Details

- **Vision Encoder**: `google/siglip2-large-patch16-384`
- **Feature Dimension**: 1024
- **Sampling Rate**: 1 frame every 16 frames
- **Normalization**: L2 Normalized (Unit Hypersphere)

---

## Contents

The dataset consists of `.txt` files corresponding to a video. Each file follows a matrix format of shape (T, D).

**Where:**
- **T** is the number of sampled temporal segments.
- **D** is the feature dimension. (1024 for `google/siglip2-large-patch16-384`)

```plaintext
0.023145 -0.012834 ... (D values)
0.018234 -0.009123 ... (D values)
```

#### Training List (`train_list.txt`)
Format: `[feature_path] [label]`
- **Sample**: `data/Normal_Videos_196_x264.txt normal`
- **Details**: Simple video-level labels for weakly supervised training.

#### Testing List (`test_list.txt`)
Format: `[feature_path] [class] [total_frames] [start1] [end1] [start2] [end2]`
- **Sample**: `data/Abuse028_x264.txt Abuse 1424 165 240 -1 -1`
- **Details**: Temporal annotations for frame-level evaluation. `total_frames` is estimated as `num_segments * 16`.

---

## How to use

These features are intended for use with the **SigMamba** training pipeline.

---

## License

Copyright &copy; 2026 Vinay Umrethe.

This dataset is available under the **Creative Commons Attribution Share Alike 4.0 International** License.