IMPACT v1.1

IMPACT is a synchronized five-view RGB-D dataset and benchmark for multi-granularity human procedural action understanding in industrial assembly. It contains 112 trials from 13 participants and 39.5 video hours across one egocentric and four exocentric views.

• Project page
• Benchmark code and task protocols
• Google Drive mirror

Release Update

July 2026, v1.1. All 560 TAS-B annotation files were revalidated, and 117 of 560 RGB videos (20.9%) were updated. Dependent task assets and selected TAS-S boundaries were regenerated from the refined canonical annotations. See v1.1/CHANGELOG.md for the complete record.

Benchmark Tasks

Family	Tasks	Metrics
Temporal Understanding	TAS-S, TAS-BL/BR	Accuracy, Edit, F1@{10,25,50}
Cross-View Understanding	CV-TA, CV-SMR, CV-SMC	Recall@{1,5}, Median Rank, Coverage, mAP, Top-1, Macro-F1
Action Forecasting	AF-S, AF-L	mR@5, AUED, ED@z
State & Reasoning	PSR, ASR, PPR-L/R, ATR-L/R	task-specific F1, delay, POS, state, phase, and multi-label metrics

Formal task definitions, split semantics, baselines, and evaluation code are versioned in the benchmark repository.

Download

The v1.1/ directory provides independent archives for annotations, RGB videos, depth streams, ego audio, eye tracking, and I3D, MViTv2, and VideoMAEv2 features. Every ZIP extracts into the same IMPACT-v1.1/ root.

The 2.8 GiB quick-start sample is available at v1.1/sample/IMPACT-v1.1-sample.zip. It contains three complete executions with synchronized media, all released feature types, and corresponding task annotations.

Selective download and automatic checksum verification:

git clone https://github.com/Kratos-Wen/IMPACT.git
cd IMPACT
python -m pip install -r requirements-tools.txt
python tools/download_impact.py --include sample --output downloads
python tools/verify_release.py downloads/v1.1/sample/IMPACT-v1.1-sample.zip

Use python tools/download_impact.py --list to inspect all bundle keys. Exact byte sizes, SHA-256 checksums, and verification scope are published in v1.1/MANIFEST.tsv, v1.1/SHA256SUMS, and v1.1/VERIFICATION.tsv.

Data Structure

IMPACT-v1.1/
├── annotations/{TAS-S,TAS-B,AF-S,ASR,PSR,PPR,ATR}/
├── videos/{ego,front,left,right,top}/
├── depth/{front,left,right,top}/
├── audio/ego/
├── eye_tracking/ego/
├── features/{I3D,MViTv2,VideoMAEv2}/
└── metadata/

The canonical identifier is <execution_id>_<view>. All views from the same trial are co-assigned in official splits to prevent cross-view leakage.

Features

Released arrays contain frozen I3D (1024-D), MViTv2-B (768-D), and VideoMAEv2-G (1408-D) representations. Extraction conventions and reference scripts are documented in the feature guide.

Ethics and Intended Use

Participants provided informed consent for data collection and research release. Public identifiers are anonymized, and the release excludes direct personal identifiers. The dataset is intended for non-commercial research and education in video understanding, procedural reasoning, forecasting, and industrial human-assistance systems. It must not be used to identify participants or infer sensitive personal attributes.

IMPACT represents a controlled industrial assembly setting and should not be treated as evidence that a model is safe for autonomous deployment. Users should assess domain shift, demographic coverage, failure modes, and human oversight requirements in their target environment. See the repository's Ethics and Intended Use statement.

Citation

The manuscript is under review at the ACM Multimedia 2026 Dataset Track. Stable versioned author metadata is available in CITATION.cff; proceedings metadata will be added after publication.

License

Dataset assets and release documentation are licensed under CC BY-NC-SA 4.0. Benchmark code is distributed separately under Apache-2.0, and third-party methods retain their respective upstream terms.