Robust Text-to-Image Person Re-Identification via Neighbor Consistency-Based Correspondence Estimation

Abstract

Text-to-image person re-identification (TIReID) retrieves target individuals using textual queries, demanding precise alignment of sparse visual-textual cues. Real-world training data, however, suffers from noisy correspondence (NC), i.e., misaligned image-text pairs from annotation errors or ambiguities. While existing NC methods work for generic retrieval, they falter in fine-grained TIReID (identity-rich but pair-limited environments); crucially, their per-sample loss filtering mistakenly discards discriminative hard samples essential for subtle identity distinctions. To address these limitations, we propose Neighbor Consistency-based Correspondence Estimation (NCCE), a unified framework that integrates cross-modal context modeling and progressive noise suppression. Specifically, the Neighbor-Consistent Noise Identification (NCNI) strategy moves beyond per-sample analysis by inferring textual descriptions from the visual neighbors of an image and evaluating the consistency between the aggregated neighbor-inferred text and the original text via weighted similarity, thus avoiding the pitfalls of loss-based filtering. Moreover, Modality-Shared Projection Learning (MSPL) addresses fine-grained mismatches by projecting both modalities into a unified space, enabling adaptive alignment through mutual regularization. Extensive experiments on three public TIReID benchmarks validate the superiority of the proposed method. Notably, on the ICFG-PEDES dataset with a 50% noise ratio, our NCCE method achieves state-of-the-art (SOTA) performance, obtaining a Rank-1 accuracy of 66.15% (+2.48% improvement) and a mAP of 40.82% (+3.51% improvement). Consistent improvements of 1-3% in both metrics are observed on the other two datasets. Furthermore, our method demonstrates strong generalizability by achieving new SOTA results on the general image retrieval benchmark MS-COCO5K.