Weakly supervised image segmentation beyond tight bounding box annotations.

Area of Science:

• Computer Vision
• Machine Learning
• Image Analysis

Background:

• Weakly supervised image segmentation typically requires precise (tight) bounding box annotations for high performance.
• Acquiring tight bounding boxes is labor-intensive and challenging compared to looser annotations.
• Existing methods struggle to maintain segmentation accuracy when using less precise bounding box supervision.

Purpose of the Study:

• To investigate the feasibility of achieving high image segmentation performance using only loose bounding box supervision.
• To develop a novel weakly supervised image segmentation approach that is robust to annotation precision.
• To extend prior multiple instance learning (MIL) techniques for improved segmentation with varied bounding box tightness.

Main Methods:

• Integration of a polar transformation-based MIL strategy with existing parallel transformation MIL for image segmentation.
• Definition of positive bags as pixels along polar lines within bounding boxes, aiding object localization.
• Introduction of a weighted smooth maximum approximation to prioritize pixels closer to the polar transformation origin.

Main Results:

• The proposed approach demonstrates state-of-the-art performance across various bounding box precision levels.
• The method shows robustness to mild and moderate inaccuracies in loose bounding box annotations.
• Evaluation on two public datasets using the Dice coefficient confirms the effectiveness of the polar transformation MIL strategy.

Conclusions:

• Weakly supervised image segmentation can achieve high performance using loose bounding boxes with the proposed polar transformation MIL method.
• The approach effectively addresses the practical challenge of acquiring precise annotations in image segmentation tasks.
• This work advances the field by enabling robust and accurate segmentation under less stringent supervision conditions.