DGRNet: A Dual-Level Graph Relation Network for Video Object Detection.

Area of Science:

• Computer Vision
• Machine Learning
• Artificial Intelligence

Background:

• Video object detection is crucial in computer vision.
• Current methods aggregate frame features but struggle with unstable feature-to-feature (Fea2Fea) relations due to occlusion, blur, or pose variations.
• This instability limits detection performance in challenging video scenarios.

Purpose of the Study:

• To propose a novel Dual-Level Graph Relation Network (DGRNet) for high-performance video object detection.
• To address the limitations of existing Fea2Fea relation estimation methods.
• To enhance feature aggregation in the temporal domain for more robust detection.

Main Methods:

• Introduced a novel Dual-Level Graph Relation Network (DGRNet).
• Leveraged residual graph convolutional networks to model Fea2Fea relations at both frame and proposal levels.
• Incorporated a node topology affinity measure to adaptively refine graph structures by pruning unreliable connections.

Main Results:

• The DGRNet method demonstrated superior performance compared to state-of-the-art methods on the ImageNet VID dataset.
• Achieved 85.0% mAP with ResNet-101 and 86.2% mAP with ResNeXt-101.
• The dual-level graph relation approach effectively improved feature aggregation and detection accuracy.

Conclusions:

• DGRNet is the first method to utilize dual-level graph relations for guiding feature aggregation in video object detection.
• The proposed approach offers a more stable and effective way to model Fea2Fea relations.
• The results confirm the effectiveness of DGRNet for enhancing video object detection capabilities.