MINet: Meta-Learning Instance Identifiers for Video Object Detection.

Area of Science:

• Computer Science
• Artificial Intelligence
• Machine Learning

Background:

• Video object detection benefits from temporal coherence across frames.
• Existing methods often require auxiliary inputs like optical flow or complex post-processing for instance association.

Purpose of the Study:

• To develop a simple yet effective method for robust online instance association in video object detection.
• To introduce a meta-learning approach that learns instance identifiers without auxiliary inputs or post-processing.

Main Methods:

• Proposed Meta-Learnt Instance Identifier Networks (MINet) that meta-learns instance identifiers.
• MINet learns instance identifier weights on-the-fly based on previous frame detections, enabling adaptation to novel frames.
• Implemented a memory mechanism for MINet to store and retrieve detection results, enhancing temporal consistency.

Main Results:

• MINet achieves robust online linking of instances within a single forward-pass.
• The method is pluggable into existing object detection models.
• Achieved 80.2% mAP on the ImageNet VID dataset when integrated with Faster R-CNN.

Conclusions:

• MINet offers a superior, pluggable solution for video object detection by effectively handling instance association.
• The meta-learning paradigm allows for flexible adaptation and robust temporal consistency.
• Demonstrated significant performance gains on the ImageNet VID dataset.