Jiacheng Xu, Fengyi Song · ACM International Conference on Multimedia (ACM MM) 2026 · 2026
DOI: 10.1145/3767308.3832547
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).
Open-vocabulary object detection (OVOD) relies on vision-language foundation models to connect visual regions with open-ended textual concepts, yet their globally aligned representations are better suited to holistic image--text alignment than fine-grained local reasoning. Recent distillation-based methods partially alleviate this mismatch by transferring cropped region features from CLIP, but they largely treat regions independently and therefore underuse the semantic-spatial context that is often crucial for subtle category discrimination. We address this issue with Object-Focused Neighborhood Distillation (OFND), which constructs object-focused neighborhood structure from patch-, instance-, and part-level regions and transfers it from a frozen CLIP teacher to a student through node-level, edge-level, and fused-representation consistency. A cross-granularity fusion module further aggregates complementary cues across heterogeneous region partitions. Experiments on OV-COCO, OV-LVIS, and Objects365 show that OFND consistently improves over direct region-feature distillation, reaching 47.3% AP50 novel on OV-COCO and 36.2% mAPr on OV-LVIS. The results indicate that modeling object-focused neighborhood structure is an effective way to adapt globally-aligned CLIP to fine-grained alignment in OVOD.
No comments yet — start the discussion below.