Self-Supervised Human Detection and Segmentation via Background Inpainting.

Area of Science:

• Computer Vision
• Machine Learning
• Artificial Intelligence

Background:

• Supervised object detection and segmentation models struggle with generalization to diverse image appearances.
• Manual data annotation is costly and time-consuming, limiting the applicability of supervised methods.
• Existing self-supervised methods have limitations in handling variations in real-world image data.

Purpose of the Study:

• To develop a self-supervised approach for object detection and segmentation that generalizes well to images with appearance variations.
• To enable accurate detection and segmentation using single images from potentially moving cameras, reducing reliance on large annotated datasets.
• To address the challenge of prohibitively expensive data annotation in computer vision tasks.

Main Methods:

• Introduced a self-supervised detection and segmentation approach leveraging the link between object segmentation and background reconstruction.
• Developed a self-supervised loss function based on the principle that background can be synthesized from surroundings, while moving objects cannot.
• Employed a proposal-based segmentation network trained with a Monte Carlo-based strategy to explore object proposals.
• Applied the method to human detection and segmentation in challenging, non-benchmark image datasets.

Main Results:

• The proposed self-supervised method demonstrates superior performance in object detection and segmentation compared to existing self-supervised techniques.
• The approach successfully generalizes to images with significant visual differences from standard training datasets.
• Effective human detection and segmentation were achieved even with images captured by a moving camera.

Conclusions:

• Self-supervised learning offers a viable alternative to supervised methods for object detection and segmentation when data annotation is a bottleneck.
• The proposed method provides a robust solution for real-world computer vision challenges involving diverse and unannotated image data.
• This work advances the field of self-supervised learning by enabling accurate object detection and segmentation in complex scenarios.