Toward Efficient Semi-Supervised Object Detection With Detection Transformer.

Area of Science:

• Computer Vision
• Machine Learning
• Artificial Intelligence

Background:

• Semi-supervised object detection (SSOD) reduces annotation costs by utilizing unlabeled data.
• Detection Transformers (DETRs) offer end-to-end object detection without Non-Maximum Suppression (NMS).
• Existing SSOD methods are not optimized for DETR architectures, leaving a research gap.

Purpose of the Study:

• To systematically investigate and improve semi-supervised learning for DETR models.
• To address challenges in DETR-based SSOD, including sensitivity to pseudo-labels and regularization difficulties.
• To propose a novel framework, Semi-DETR++, for efficient and effective SSOD with DETRs.

Main Methods:

• Introduced a stage-wise hybrid matching strategy combining one-to-many and one-to-one assignments for pseudo-label robustness.
• Developed a re-decode query consistency training method to regularize the DETR decoder based on its layer-wise behavior.
• Preserved NMS-free inference inherent to DETR models.

Main Results:

• Semi-DETR++ demonstrates more efficient semi-supervised learning across various DETR architectures.
• The proposed framework significantly outperforms existing SSOD methods.
• Components of Semi-DETR++ show versatility and generalize well to semi-supervised segmentation tasks.

Conclusions:

• Semi-DETR++ effectively tackles key challenges in semi-supervised DETR training.
• The novel methods enhance robustness to pseudo-labels and improve decoder regularization.
• The framework offers a scalable and high-performing solution for SSOD, with potential applications in related tasks.