Semi-Supervised Object Detection: A Survey on Progress from CNN to Transformer

Area of Science:

• Computer Vision
• Machine Learning
• Artificial Intelligence

Background:

• Semi-supervised object detection (SSOD) combines limited labeled data with extensive unlabeled data.
• This approach mitigates the high cost and time associated with acquiring large labeled datasets.
• Early SSOD models struggled with effectively utilizing unlabeled data and handling noisy pseudo-labels.

Purpose of the Study:

• To provide a comprehensive review of 28 recent advancements in SSOD methodologies.
• To analyze the integration of semi-supervised learning principles into object detection frameworks.
• To stimulate further research in SSOD by identifying challenges and future directions.

Main Methods:

• Review of methodologies spanning Convolutional Neural Networks (CNNs) to Transformers.
• Analysis of core semi-supervised learning components: data augmentation, pseudo-labeling, consistency regularization, and adversarial training.
• Comparative evaluation of different SSOD models based on performance and architecture.

Main Results:

• Significant improvements in SSOD performance have been achieved through recent methodological advancements.
• Effective strategies for leveraging unlabeled data and managing pseudo-label noise have been developed.
• A wide range of SSOD techniques, from CNN-based to Transformer-based, are now available.

Conclusions:

• SSOD is a rapidly evolving field with substantial progress in overcoming initial limitations.
• Further research is needed to address remaining challenges and explore novel approaches in SSOD.
• This review provides a valuable resource for researchers and practitioners in computer vision and machine learning.