| Abstract [eng] |
This thesis presents a systematic empirical investigation of the MACE (Modality-Aware Collaborative Ensemble Learning) framework for cross-modality infrared-to-visible person re-identification on the SYSU-MM01 benchmark. The study is structured in two experimental phases. The first phase examines whether standard architectural modifications, including alternative backbone networks, explicit attention mechanisms, and advanced feature fusion strategies, produce performance improvements when applied to an already well-optimized ensemble model. Results indicate that the tested modifications offered no consistent retrieval improvement over the ResNet-50 baseline, and in several cases reduced accuracy, suggesting that MACE's collaborative ensemble and knowledge distillation design is already well-calibrated for the modifications considered. The second phase investigates operationally relevant conditions absent from the original evaluation, specifically multi-query inference aggregation, training data efficiency, and robustness to partial occlusion. Multi-query average fusion substantially improves Rank-1 accuracy without any model retraining, revealing that single-query benchmarks underestimate practical retrieval capability. Training data experiments show that performance degrades gradually until approximately 50% of training identities are available, below which the ensemble's core learning mechanisms begin to fail due to insufficient identity diversity. Occlusion experiments demonstrate that the baseline model is sensitive to partial query occlusion, with the degree of degradation depending on which body region is obscured, and that occlusion-augmented training reduces this vulnerability while simultaneously improving clean performance. Together, the findings provide a comprehensive characterization of MACE's strengths, limitations, and practical operating boundaries. |
