Object Detection in Videos by High Quality Object Linking.

Area of Science:

• Computer Vision
• Machine Learning
• Artificial Intelligence

Background:

• Object detection in videos is more challenging than in static images due to quality degradation.
• Exploiting temporal contexts by linking objects across frames (tubelets) improves video object detection.
• Previous methods linking objects between neighboring frames have limitations.

Purpose of the Study:

• To improve object linking quality for better video object detection and classification.
• To propose a novel method for linking objects within the same frame, unlike previous frame-to-frame approaches.
• To enhance the aggregation of classification scores using high-quality object tubelets.

Main Methods:

• A cuboid proposal network was developed to extract spatio-temporal candidate cuboids for object movement.
• A short tubelet detection network was implemented to identify short object trajectories in video segments.
• A short tubelet linking algorithm was designed to connect overlapping tubelets into longer ones.

Main Results:

• The proposed method demonstrated superior performance on the ImageNet VID dataset compared to static image detectors.
• It outperformed previous state-of-the-art methods in video object detection.
• Significant improvements were observed for fast-moving objects, with an 8.8% increase over static image detectors.

Conclusions:

• The novel approach of within-frame object linking enhances temporal context utilization for video object detection.
• The method effectively addresses challenges posed by degraded image quality in videos.
• This work sets a new benchmark for video object detection, particularly for dynamic objects.