Do More With Less: Capacity-Aware Selective Hashing for Continual Cross-Modal Retrieval

Although cross-modal hashing enables efficient large-scale retrieval by encoding multimodal data into compact binary representations, its fixed code length and binary nature impose a fundamental capacity constraint that hinders continual adaptation to growing data streams and emerging semantic concepts. Existing continual cross-modal hashing methods typically resort to re-indexing or code expansion to accommodate new tasks, which either incur prohibitive computational costs or disrupt the consistency of the established Hamming space. More fundamentally, under a fixed bit budget, the continual accumulation of semantic information inevitably saturates the limited representation capacity, leading to intensified bit collisions and degraded neighborhood structures, and thereby exacerbating the stability-plasticity conflict that limits long-term retrieval performance. To address this, we propose Capacity-Aware Selective Hashing (CASH), which significantly improves Hamming-space utilization through bit-level selective allocation under a fixed capacity budget, enabling stable continual learning while preserving long-term code compatibility. CASH employs a coarse-fine dual-branch hashing network to provide complementary global and fine-grained code candidates, and introduces a task-prompt-conditioned bit-selection mechanism that dynamically assigns each bit to the branch with the higher discriminative utility, effectively mitigating bit collisions and cross-task interference. To further ensure stability, the multimodal encoders are frozen, and incremental adaptation is achieved via lightweight task prompts. Extensive experiments under standard incremental protocols demonstrate that our CASH consistently outperforms SOTA baselines in both retrieval accuracy and long-term stability across task partition schemes.