- Rating
- 2
- Confidence
- 3
- Informativeness
- 2
- Presentation
- 2
- Soundness
- 2
Read review
Paper summary
To address the problem that pretrained visual models tend to rank visually similar distractors ahead of true matches, the authors propose the SAFR framework. During training, the SAFR framework introduces Foundation-Error-Aware Ranking, which uses ranking errors from a frozen foundation model to dynamically adjust the margin constraints on hard samples, thereby improving the model's ability to distinguish identities. During retrieval, it employs Sparse Foundation-Affinity Correction, which combines global and local feature affinities to correct the final matching scores. Experiments on RGBNT201, RGBNT100, and MSVR310 show that SAFR outperforms the comparison methods listed in the tables.
Critical strengths
- Section 3.1 proposes a method that uses incorrect rankings of positive and negative samples from a frozen foundation model to adjust the training margins, rather than directly learning the foundation model's similarity relations. I think using these ranking errors to improve the model's ability to distinguish identities is a good idea.
- Table 4 in Section 4.4 compares different retrieval score correction methods. These detailed comparisons help better understand the actual impact of affinity correction on retrieval results and demonstrate the authors' solid experimental work.
Critical weaknesses
- The main ablation experiments in Section 4.4 are all conducted on RGBNT201, without corresponding analyses on the two vehicle Re-ID datasets, RGBNT100 and MSVR310.
- In Table 5 of Section 4.4, using only global affinity already achieves 83.5% mAP, while adding local affinity improves it only to 83.7%, with no change in Rank-1. This result makes me skeptical about the actual contribution of local features.
Questions for authors
- Regarding the foundation model's ranking reversal magnitude in Eq. (7), does the performance improvement actually come from the error information provided by the foundation model, or simply from imposing stronger ranking constraints on hard samples?
- The main ablation studies in Section 4.4 are limited to RGBNT201. Do the two proposed components provide similar improvements on RGBNT100 and MSVR310?