Attention-Guided Disentangled Feature Aggregation for Video Object Detection.

Area of Science:

• Computer Vision
• Artificial Intelligence
• Machine Learning

Background:

• Object detection in still images is well-established, but video object detection faces challenges like blur and occlusion.
• Existing methods struggle with the dynamic nature of video data, impacting localization and classification accuracy.

Purpose of the Study:

• To propose an attention-heavy framework for robust video object detection.
• To address challenges in video object detection by effectively aggregating frame-level features.
• To improve the performance of object detection in video sequences.

Main Methods:

• A two-stage object detection framework based on Faster R-CNN architecture.
• Integration of scale, spatial, and task-aware attention in a disentanglement head.
• Utilization of temporal attention in an aggregation head to combine features from support frames.

Main Results:

• Achieved a mean Average Precision (mAP) of 49.8 with ResNet-50 backbone.
• Achieved a mean Average Precision (mAP) of 52.5 with ResNet-101 backbone on the ImageNet VID dataset.
• Demonstrated significant performance improvement over individual baseline methods.

Conclusions:

• The proposed attention-heavy framework effectively tackles challenges in video object detection.
• Feature disentanglement and temporal aggregation significantly enhance detection accuracy.
• The approach offers a promising solution for real-world video analysis applications.