Yandan Wang, Jiangyan Chen, Yinglong Ma, Chenqi Guo, Jiaqi Ma, Weiming Dong · ACM International Conference on Multimedia (ACM MM) 2026 · 2026
DOI: 10.1145/3767308.3834955
Counts differ because each database indexes a different set of publications. We treat OpenAlex as the canonical count; Google Scholar is not shown (no API, and crawling it violates its ToS).
Query-based object detectors typically employ one-to-many matching mechanisms to enhance training supervision. However, in long-tailed scenarios, such methods are hindered by unreliable query representations and rigid positive sample assignment strategies, leading to imbalanced supervision intensity across different classes. In this paper, we present an object-level analysis of the interplay between positive sample quantity and matching quality, revealing significant disparities in supervision intensity among classes of varying frequencies. Motivated by these insights, we propose SIGMA (Supervision-Intensity Guided one-to-many MAtching), a novel framework for long-tailed object detection. Specifically, we design an adaptive matching strategy that constructs positive sample sets based on the supervision intensity of each object, effectively enriching the training supervision, particularly for objects of rare classes. Furthermore, we introduce Semantic-Aware Query Initialization and Region-Aligned Query Refinement to enhance query quality for tail classes and ensure reliable matching. Extensive experiments on LVIS v1.0 and COCO-LT demonstrate that SIGMA achieves state-of-the-art performance, with significant improvements in tail class detection.
No comments yet — start the discussion below.