Benefiting From OOD Samples in Open-Set Semi-Supervised Object Detection

Area of Science:

• Computer Vision
• Machine Learning
• Artificial Intelligence

Background:

• Open-set semi-supervised object detection (OSSOD) addresses challenges where unlabeled data may contain both in-distribution (ID) and out-of-distribution (OOD) samples.
• Existing OSSOD methods typically aim to filter out OOD samples, potentially losing valuable information for feature learning.

Purpose of the Study:

• To develop a novel approach for OSSOD that leverages OOD samples to enhance the detection of ID categories.
• To improve feature learning and overall detection performance in open-set conditions by effectively utilizing all unlabeled data.

Main Methods:

• Instance-level consistency regularization (ICR) applied to all detected instances, including OOD samples.
• OOD-aware contrastive learning (OCL) to cluster ID objects and separate OOD samples in feature space.
• Prototype-based multimetric adaptive matching (MAM) for reliable ID/OOD sample identification and weighted consistency regularization.

Main Results:

• The proposed method effectively utilizes OOD samples to optimize feature learning without negative impacts on semi-supervised learning.
• Significant improvements in detection capability for ID categories in open-set scenarios were achieved.
• The approach demonstrated superior performance compared to current state-of-the-art OSSOD methods.

Conclusions:

• Leveraging OOD samples in unlabeled data can boost feature learning and detection performance in OSSOD.
• The proposed OCL and MAM methods provide effective mechanisms for distinguishing and utilizing ID and OOD samples.
• This work offers a promising direction for advancing OSSOD by embracing, rather than excluding, OOD data.