KratosWen/IMPACT

---
pretty_name: IMPACT
license: other
task_categories:
- video-classification
- image-classification
- audio-classification
- text-classification
language:
- en
tags:
- industrial-assembly
- procedural-understanding
- action-segmentation
- action-forecasting
- reasoning
- multi-view
size_categories:
- 100K<n<1M
---

# IMPACT

IMPACT is a multi-view industrial assembly dataset and benchmark for fine-grained action understanding, procedural reasoning, action forecasting, and anomaly-aware evaluation.

- Project page: https://kratos-wen.github.io/IMPACT/
- Repository: https://github.com/Kratos-Wen/IMPACT
- Google Drive release: https://drive.google.com/drive/folders/1P7vBnxSVH9g_lQc5n0c0WA47QAGEhE_U?usp=sharing

## Release Layout

This Hugging Face dataset mirrors the public IMPACT release bundles:

- `annotations/annotations_v1.zip`: official annotation release bundle
- `videos/videos_{ego,front,left,right,top}.zip`: five-view RGB archives
- `depth/depth_{front,left,right,top}.zip`: exocentric depth archives
- `audio/audio_ego.zip`: egocentric audio bundle
- `eye_tracking/eye_tracking_ego.zip`: egocentric eye-tracking TSV bundle
- `features/features_{I3D,MViTv2,VideoMAEv2}.zip`: released feature bundles
- `sample/`: quick-start subset with 3 executions, multi-view media, eye-tracking traces, released features, and task annotations

## Benchmark Coverage

The current public code release provides runnable benchmark wrappers for:

- `TAS`
- `CV-TA`
- `CV-SM`
- `AF-S`
- `PSR`
- `ASR`
- `PPR`
- `ATR`

The repository also documents `AF-L` with its task protocol and reported baseline design.

## Sample Subset

The sample subset is intended for quick-start inspection, reviewer access, and lightweight debugging.

It currently contains:

- 3 executions
- 15 RGB video files
- 12 depth files
- 3 ego-audio files
- 3 ego eye-tracking files
- 45 feature files
- task-specific annotation folders covering `TAS-S`, `TAS-B`, `ASR`, `PSR`, `PPR`, and `ATR`

Selected sample executions:

- `AL07EJ17_Disassembly_A_001`
- `AL07EJ17_Reassembly_A_002`
- `ER07AD15_Disassembly_A_001`

## Notes

- The GitHub repository stores lightweight protocol assets under `dataset/`; larger release bundles are mirrored here and on Google Drive.
- The Hugging Face release does not include internal work-in-progress annotation revisions.
- License details for repository-authored code and dataset assets are documented in the GitHub repository.