Semantic Hardness Is Not Visual Hardness: Sign-Aware Hard Negative Mining for Sign Language Retrieval
For sign language retrieval researchers, this work addresses a known bottleneck in fine-grained retrieval by introducing a visually-aware negative mining strategy, though it is an incremental improvement over existing hard negative mining methods.
The paper identifies that fine-grained sign language retrieval failures are due to ineffective hard negative supervision, not model capacity. It proposes Sign-Aware Hard Negative Mining (SAN), which selects hard negatives based on visual confusability, achieving substantial improvements on PHOENIX-2014T while preserving coarse-grained accuracy.
Sign Language Retrieval (SLRet) enables efficient access to sign language content but remains fragile in fine-grained scenarios where visually similar signs must be distinguished. We show that this limitation does not stem from model capacity, but from ineffective hard negative supervision. Specifically, we formulate fine-grained retrieval failures as a negative distribution mismatch: semantically distinct yet visually confusable signs are rarely treated as hard negatives, while existing text-based mining strategies fail to capture such visual ambiguity. To address this issue, we propose Sign-Aware Hard Negative Mining (SAN), which constructs hard negatives based on visual confusability in the sign embedding space rather than linguistic similarity. Experiments on PHOENIX-2014T demonstrate that SAN substantially improves fine-grained retrieval performance while preserving coarse-grained accuracy, highlighting the importance of aligning negative supervision with visual ambiguity in sign language retrieval.