Deployment Prior Injection for Run-Time Re-Biasable Object Detection.

Area of Science:

• Computer Vision
• Machine Learning
• Artificial Intelligence

Background:

• Object relation context improves object detection when training and testing data distributions align.
• Shifting data distributions across space and time introduce harmful training set bias in object detectors.
• Existing detectors lack the ability to incorporate deployment context priors during testing without parameter updates.

Purpose of the Study:

• To develop an object detector capable of incorporating deployment context priors at runtime without parameter updates.
• To enable detectors to explicitly learn disentangled representations with respect to context priors.
• To introduce a method for 're-biasing' detectors towards a given context prior dynamically.

Main Methods:

• Introduced an additional graph input representing deployment context prior, with edge values signifying object relations.
• Trained the detector with a modified objective to ensure its behavior is constrained by the graph input.
• Enabled runtime adaptation by allowing graph edits to inject deployment context priors without parameter updates.

Main Results:

• The proposed detector can be re-biased at runtime using graph edits to adapt to specific deployment contexts.
• The detector demonstrates self-re-biasing capability using approximated deployment priors when the actual prior is unknown.
• Experiments on COCO and Objects365 datasets confirm the effectiveness of the run-time re-biasable detector.

Conclusions:

• The developed detector effectively addresses the challenge of shifting data distributions by enabling runtime context adaptation.
• This approach allows for flexible and efficient deployment of object detectors in diverse and dynamic environments.
• The ability to re-bias detectors without parameter updates significantly enhances their practical applicability and robustness.