CUHK-S is a privacy-preserving subset of the CUHK-X dataset, a large-scale multimodal benchmark for Human Action Recognition, Understanding, and Reasoning. It removes RGB video to prevent facial identification and downscales visual modalities to 320 × 240. The dataset was accepted at MobiSys 2026 and is hosted by willow-cuhk.
Use Cases
- Training models for human action recognition based on non-RGB visual modalities.
- Benchmarking privacy-preserving computer vision systems based on downscaled visual data.
- Developing multimodal reasoning systems for human activities based on the dataset's multimodal benchmark.
- Studying human action understanding in scenarios where facial identification is prohibited.
Strengths
- Designed for privacy by removing RGB video to prevent facial identification.
- Visual modalities are downscaled to a consistent resolution of 320 × 240.
- Part of a large-scale multimodal benchmark accepted at a major conference (MobiSys 2026).
Limitations
- Description metadata is limited; actual data quality requires manual inspection after download.
- Column-level documentation is absent; field semantics must be inferred after download.
- Row count is unknown, which may limit suitability assessment.
Provenance
- Source
- Subset of the CUHK-X dataset, hosted by willow-cuhk.
- Freshness
- Last updated 2026-05-03 03:39:51; freshness should be verified.