Dense Information Learning based Semi-Supervised Object Detection.

Area of Science:

• Computer Science
• Artificial Intelligence
• Machine Learning

Background:

• Semi-Supervised Object Detection (SSOD) utilizes unlabeled data but current methods are passive and overlook category relationships.
• Existing SSOD techniques often rely solely on original image data and prioritize teacher model predictions over inter-category dependencies.

Purpose of the Study:

• To introduce Dense Information Learning (DIL), an active approach for SSOD that generates densely informative unlabeled data.
• To enhance the utilization of unlabeled data by actively incorporating exploitable information and enforcing relation consistency.

Main Methods:

• Dense Information Augmentation (DIA): Actively generates unlabeled data with exploitable information using a foreground bank, enhancing and filtering noise.
• Relation Consistency Regularization (RCR): Enforces manifold-level consistency under feature and image perturbations, guiding the network to focus on discriminative features.

Main Results:

• DIL significantly improves semi-supervised object detection performance by effectively leveraging unlabeled data.
• On the MS-COCO dataset, DIL achieved relative improvements of 12.6% and 10.0% in mAP using 5% and 10% labeled data, respectively, compared to supervised baselines.

Conclusions:

• The proposed DIL approach demonstrates a novel and effective way to actively learn from unlabeled data in SSOD.
• DIL's active data generation and relation consistency mechanisms lead to substantial performance gains, outperforming passive methods.