Connecting the objects: Relational occlusion graphs for amodal segmentation

  • Liu, Huiling; 
  • Cho, Eunnam; 
  • Choi, Soyun; 
  • Jung, Aecheon; 
  • Kim, Seung Wook; 
  • ... Hong, Sungeun
Citations

WEB OF SCIENCE

0
Citations

SCOPUS

0

초록

Amodal instance segmentation predicts complete object masks by inferring occluded regions. Existing methods typically rely on object-centric cues like shape priors or intra-object coherence. We argue that relationships between neighboring objects, such as relative position, overlap, and occlusion patterns, play a crucial role in inferring missing geometries. We propose Amodal Relational Occlusion Graph (AROG) to explicitly model inter-object occlusion. Our framework dynamically constructs a directed and weighted graph from evolving object queries. Each node in this graph represents an object query. Each directed edge encodes the specific occlusion influence from one object to another. We integrate this graph into the self-attention mechanism as a relational inductive bias. This design enables occluded queries to attend selectively to relevant occluders. AROG achieves strong performance across multiple benchmarks and shows that structured relational modeling provides an effective perspective for amodal perception.

키워드

Amodal instance segmentation; Occlusion modeling; Relational graphs
제목
Connecting the objects: Relational occlusion graphs for amodal segmentation
저자
Liu, Huiling; Cho, Eunnam; Choi, Soyun; Jung, Aecheon; Kim, Seung Wook; Hong, Sungeun
DOI
10.1016/j.patrec.2026.06.021
발행일
2026-09
유형
Article
저널명
Pattern Recognition Letters
권
207
페이지
163 ~ 170