Researchers present a parameter-efficient CLIP adaptation framework for long-term animal re-identification that addresses challenges posed by gradual morphological evolution and seasonal appearance shifts. The method introduces a continuous metadata-conditioning mechanism that incorporates numerical attributes directly into the prompt representation during training, preserving their structure rather than discretizing them.
- The framework combines low-rank visual adaptation, prompt-based supervision, and cross-modal alignment to adapt CLIP for ecological settings.
- Continuous metadata conditioning enables smooth modulation of the embedding space while maintaining a purely visual inference pipeline at test time.
- Experiments on a seven-year longitudinal fish dataset and multiple wildlife benchmarks show improved performance under closed-set, open-set, and time-aware evaluation protocols.
The approach improves robustness to longitudinal appearance variation and temporal distribution shifts without requiring metadata during inference.