Motion Representations for Privacy-Aware Cross-Domain Action Recognition

A 342.8 KB PDF authored by Pascal Benschop, last updated on 2026-06-03, describes a study on action recognition using motion-focused representations. The work evaluates motion-history images and optical flow on datasets including HMDB-51 and UCF-101 to reduce identifiable appearance cues while retaining action recognition accuracy.

Use Cases

• Training privacy-preserving action recognition models based on motion-history images and optical flow representations described in the study.
• Benchmarking zero-shot and few-shot action recognition performance on HMDB-51 and UCF-101 datasets using motion-only inputs.
• Evaluating the reduction of sensitive attribute predictability, such as skin color, in video data as discussed in the description.

Strengths

• The study reports specific performance metrics, including 31% and 52% zero-shot top-1 accuracy on HMDB-51 and UCF-101, respectively.
• Results indicate optical flow representation approached chance level for skin-color prediction on PA-HMDB51, suggesting a privacy benefit.
• The dataset is shared under a CC-BY-4.0 license, permitting reuse with attribution.

Limitations

• The actual data files (e.g., video clips or motion representations) are not included; the provided file is a 342.8 KB PDF describing the research.
• Column-level documentation and sample data are unavailable, making direct data inspection impossible.
• The scale of the underlying video data used in the study is unknown.

Provenance

Source

Pascal Benschop via figshare

Collection Method

Likely derived from established action recognition datasets (HMDB-51, UCF-101, Kinetics-400) processed to generate motion representations.

Freshness

Last updated 2026-06-03 06:04:51.