EmbedTPR: Reliability-Aware Reconstruction and Query-Adaptive Fusion for Open-Vocabulary Text-Based Person Retrieval

Open-vocabulary text-based person retrieval must match natural-language descriptions containing concepts that may be absent from task-specific training data while controlling unreliable generated evidence and inference cost. This paper formulates heterogeneous evidence reconstruction and masked branch fusion as a trainable retrieval procedure around frozen foundation backbones. EmbedTPR contains two interconnected learnable modules. Unified Heterogeneous Evidence Reconstruction (UHER) aligns gallery-side generated semantics with visual evidence and estimates attribute reliability; for a text-only query, the semantic path operates without visual tokens and uses generation agreement to calibrate query attributes. Query-Adaptive Coherent Evidence Fusion (QACEF) predicts query-specific weights over the three branches available to a text query through a retrieval–consistency–calibration objective. The optional image-to-image branch is part of the general formulation but is not evaluated or claimed in this study. Only 15.2 M of 5015 M parameters are trainable. EmbedTPR obtains 49.66% Rank-1 and 64.45% mAP on CUHK-PEDES, 46.20% Rank-1 and 58.30% mAP on ICFG-PEDES, and 54.30% Rank-1 and 68.10% mAP on RSTPReid. Its CUHK-PEDES Rank-1 is lower than that of CLIP ViT-L/14, whereas Rank-5, Rank-10, and mAP are higher. Cross-dataset transfer from CUHK-PEDES to ICFG-PEDES reaches 18.20% Rank-1. On the sequestered FineGrained-1282 query/label evaluation manifest, learnable fusion improves mAP by 4.04 points over fixed weights. Component removal, evidence perturbation, and 1k–100k gallery tests characterize the evaluated accuracy, robustness, and computational scope. These results support reliability-calibrated query-adaptive fusion for text-query person retrieval; broader multimodal-query and open-world generalization remain to be validated. The source code, configurations, prompts, and evaluation scripts will be released publicly upon acceptance.