AiWo presents at IEEE Big Data 2025: Revisiting domain similarity for cross‑domain few‑shot activity recognition
We’re delighted to share that one of AiWo’s doctoral researchers, Changyi Li presents his poster titled “Revisiting the Impact of Domain Similarity on the Performance of Cross-Domain Few-Shot Activity Recognition” at the 13th IEEE International Conference on Big Data (IEEE Big Data 2025).
- Paper page: https://research.aalto.fi/en/publications/revisiting-the-impact-of-domain-similarity-on-the-performance-of-/
- Conference: https://conferences.cis.um.edu.mo/ieeebigdata2025/
- Date: 8 Dec 2025 – 11 Dec 2025
About the research
Video-based activity recognition is widely used in industrial safety, manufacturing, smart cities, and human–robot collaboration. A persistent challenge is that target domains (e.g., a specific factory) often have very few labeled videos, making it difficult to train reliable models. Cross-Domain Few-Shot Learning (CDFSL) addresses this by transferring knowledge from a label-rich source dataset to a data-scarce target domain.
This work revisits a key but underexplored question: how do similarities between source and target domains affect recognition performance in the target domain? Existing similarity metrics (e.g., Maximum Mean Discrepancy) mostly compare distributions and often fall short in guiding practitioners toward effective source dataset choices.
Key ideas and contributions
- Moves beyond generic statistical metrics to examine practical, domain-level attributes such as camera angle, background scene, and label granularity.
- Systematically evaluates state-of-the-art CDFSL methods across diverse source–target pairs.
- Uses widely adopted video datasets (Kinetics‑100, HMDB51, HA‑VID, and Meccano) to ground the analysis.
- Provides actionable insights for selecting source domains that better transfer to industrial target environments.
Why it matters
In industry-facing deployments, choosing the right source data can make or break a few-shot transfer. By clarifying which aspects of domain similarity matter most, this study offers practical guidance to engineers and researchers seeking robust video activity recognition in data-scarce settings—especially in specialized industrial contexts where labeled data is limited.
Authors
- Changyi Li, Mobile Cloud Computing, Department of Information and Communications Engineering
- Yu Xiao, Mobile Cloud Computing, Department of Information and Communications Engineering
Congratulations to Changyi and the team on this acceptance!
