Delving Deeper Into Pixel Prior for Box-Supervised Semantic Segmentation.

Area of Science:

• Computer Vision
• Machine Learning
• Image Segmentation

Background:

• Weakly supervised semantic segmentation (WSSS) uses bounding boxes for object annotation.
• Existing WSSS methods often fail to fully exploit bounding box prior information.
• This limitation hinders performance, particularly for fine object parts and boundaries.

Purpose of the Study:

• To propose a novel Pixel-as-Instance Prior (PIP) for WSSS.
• To enhance the exploitation of pixel-level prior information from bounding box annotations.
• To improve segmentation accuracy for fine details and boundaries in WSSS.

Main Methods:

• Developed Pixel-as-Instance Prior (PIP) based on irregular-filling and label-ambiguity priors.
• Introduced a constrained loss similar to multiple instance learning (MIL).
• Incorporated a labeling-balance loss for joint training of WSSS models.
• Treated each pixel as a weighted positive or negative instance.

Main Results:

• PIP significantly improves the performance of various WSSS methods.
• Achieved competitive results on PASCAL VOC 2012 and Cityscapes benchmarks.
• Demonstrated negligible computational overhead during the training stage.

Conclusions:

• The proposed PIP effectively utilizes pixel prior information from bounding boxes.
• PIP enhances WSSS performance by addressing limitations in exploiting bounding box annotations.
• This approach offers a flexible and efficient way to improve WSSS accuracy.