Group Visual Relation Detection

Area of Science:

• Computer Vision
• Artificial Intelligence
• Machine Learning

Background:

• Existing visual relation detection focuses on individual objects, overlooking group interactions.
• Groups are prevalent in image semantics, necessitating advanced detection methods.
• Group Visual Relation Detection (GVRD) extends traditional methods by including group subjects and/or objects.

Purpose of the Study:

• To introduce a novel task, Group Visual Relation Detection (GVRD).
• To propose a new method, Simultaneous Group Relation Prediction (SGRP), for addressing GVRD.
• To create and release a new dataset, COCO-GVR, for GVRD research.

Main Methods:

• The Simultaneous Group Relation Prediction (SGRP) method is proposed.
• SGRP comprises Entity Construction (EC), Feature Extraction (FE), and Group Relation Prediction (GRP) modules.
• The EC module generates instances and group/phrase candidates; FE extracts multi-modal features; GRP predicts groups and predicates simultaneously.

Main Results:

• The novel COCO-GVR dataset, containing 9,570 images and 31,855 GVRs, was created.
• Extensive experiments were conducted on the COCO-GVR dataset.
• The SGRP method demonstrated superior performance compared to baseline methods.

Conclusions:

• GVRD is a valuable extension to visual relation detection.
• The SGRP method effectively addresses the GVRD task.
• The COCO-GVR dataset facilitates future research in group visual relation detection.