AIDA-ReID: Adaptive Intermediate Domain Adaptation for Person Re-Identification
A new approach called Adaptive Intermediate Domain Adaptation (AIDA), or Source-Free Multi-Source Intermediate Domain Adaptation (SF-MIDA), has been developed to handle domain shifts in person re-identification (Re-ID). Re-ID aims to recognize people across various camera angles, but it often struggles due to issues like changes in lighting, different backgrounds, and camera variations. Existing methods like IDM and IDM++ rely on fixed techniques and need both source and target data at the same time, which limits their use in multi-source or source-free scenarios. AIDA overcomes these limitations by treating the problem as a source-free multi-source challenge, enabling adaptation to new settings without accessing target data. You can find more details in a paper on arXiv (2605.00111).
Key facts
- AIDA is also referred to as Source-Free Multi-Source Intermediate Domain Adaptation (SF-MIDA).
- Person Re-ID matches individuals across non-overlapping camera views.
- Domain shifts arise from variations in illumination, background, camera characteristics, and population distributions.
- Supervised models degrade in unseen environments.
- IDM and IDM++ are existing intermediate domain approaches with fixed mixing strategies and joint source-target access.
- AIDA addresses limitations of IDM and IDM++ in multi-source and source-free settings.
- The paper is published on arXiv with ID 2605.00111.
- AIDA does not require access to target domain data.
Entities
Institutions
- arXiv