ART-VOD: Adaptive Relational Distillation and Temporal Modeling for Open-Vocabulary Video Object Detection

Abstract

Open-vocabulary video object detection aims to recognize and localize objects in videos from an unrestricted set of categories. Despite recent advances driven by knowledge transfer from large-scale vision–language models, existing methods still suffer from insufficient relational modeling, unreliable similarity matching, and inaccurate feature aggregation, which collectively hinder discriminative semantic learning and temporal coherence. To address these challenges systematically, we propose ART-VOD, an adaptive relational distillation and temporal modeling method designed to strengthen representation learning, association robustness, and temporal consistency. Specifically, we introduce an Adaptive Relational Contrastive Distillation (ARCD) module, which integrates information entropy weighting, curricular negative sample modulation, and relational contrastive loss to capture rich contextual dependencies during feature distillation, thereby improving knowledge transfer and enhancing discriminative capacity. In addition, we develop a Current Distribution-Aligned Matching (CDAM) module that establishes cross-frame correspondences by aligning the matching process with the embedding distribution of the current frame, effectively mitigating distributional drift and boosting association accuracy. Finally, we present a Semantic Relevance-Based Feature Aggregation (SRFA) module that dynamically aggregates the matched region embeddings by jointly considering frame-level reliability and region-level quality. This selective aggregation suppresses irrelevant information and enriches temporal representations. Extensive experiments on multiple datasets validate the effectiveness of our method. Notably, the proposed ART-VOD surpasses OV-VOD by 3.9% in $\text{AP}_{n}$ on the LV-VIS dataset, demonstrating its superiority over existing methods.