TIB: Detecting Unknown Objects via Two-Stream Information Bottleneck.

Area of Science:

• Computer Science
• Artificial Intelligence
• Machine Learning

Background:

• Safe application of object detectors requires detecting novel objects unseen during training.
• Unsupervised out-of-distribution object detection (OOD-OD) aims to identify unknown objects without auxiliary datasets.
• Leveraging in-distribution data is crucial for improving model discrimination in OOD-OD tasks.

Purpose of the Study:

• To propose a novel method for unsupervised out-of-distribution object detection.
• To address the challenge of detecting unknown objects without supervision.
• To enhance the discrimination capabilities of object detection models.

Main Methods:

• A Two-Stream Information Bottleneck (TIB) approach is proposed, combining a standard Information Bottleneck (IB) and a Reverse Information Bottleneck (RIB).
• The standard IB disentangles instance representations for object localization and recognition.
• RIB generates simulated out-of-distribution features to mitigate the lack of unknown data supervision, by reversing the IB optimization objective.

Main Results:

• The proposed TIB method demonstrates superior performance on OOD-OD tasks.
• Effectiveness is validated across various object detection benchmarks, including open-vocabulary, incremental, and open-set object detection.
• The mixture of information bottlenecks enhances the capture of object-related information, improving discrimination.

Conclusions:

• The TIB method offers a robust solution for unsupervised OOD-OD by effectively utilizing in-distribution data.
• The approach successfully alleviates the impact of missing supervision for unknown objects.
• This work advances the capability of object detectors to handle novel and diverse objects in real-world scenarios.