DGC-Net: Dynamic Graph Contrastive Network for Video Object Detection.

Area of Science:

• Computer Vision
• Machine Learning
• Artificial Intelligence

Background:

• Video object detection faces challenges with object appearance degradation, unlike in static images.
• Existing methods aggregate multi-frame features but neglect supervision knowledge, leading to insufficient feature aggregation and false detections.

Purpose of the Study:

• To propose a novel Dynamic Graph Contrastive Network (DGC-Net) for enhanced video object detection.
• To improve feature aggregation by incorporating supervision knowledge and addressing limitations of current methods.

Main Methods:

• Designed a frame-level graph contrastive module for aggregating frame features and exploiting contextual representations.
• Developed a proposal-level graph contrastive module for aggregating proposal features and learning semantic representations.
• Introduced a graph transformer for dynamic graph structure adjustment, pruning useless nodes/edges to reduce ambiguity and scale.

Main Results:

• DGC-Net demonstrates superior performance over state-of-the-art methods on the ImageNet VID dataset.
• Achieved 86.3% mAP with ResNet-101 and 87.3% mAP with ResNeXt-101.
• Introduced DGC-Net Lite for real-time video object detection with significantly faster inference speeds.

Conclusions:

• The proposed DGC-Net effectively addresses appearance degradation and false detection issues in video object detection.
• The dynamic graph contrastive approach enhances both contextual and semantic feature representations.
• DGC-Net offers a promising solution for accurate and efficient video object detection, with a real-time variant available.