Saliency as Pseudo-Pixel Supervision for Weakly and Semi-Supervised Semantic Segmentation.

Area of Science:

• Computer Vision
• Machine Learning

Background:

• Weakly supervised semantic segmentation faces challenges like sparse object coverage and inaccurate boundaries.
• Existing methods struggle with co-occurring pixels from non-target objects.

Purpose of the Study:

• To introduce an improved framework, Explicit Pseudo-pixel Supervision (EPS++), for semantic segmentation using image-level weak supervision.
• To enhance object boundary accuracy and reduce noise in pseudo-masks.

Main Methods:

• EPS++ combines image-level labels for object identity and saliency maps for boundary information.
• A joint training strategy leverages complementary information from both supervision types.
• An Inconsistent Region Drop (IRD) strategy effectively handles saliency map errors with fewer hyperparameters.

Main Results:

• EPS++ achieves accurate object boundaries and effectively discards co-occurring pixels, significantly improving pseudo-mask quality.
• The method establishes new state-of-the-art performances on three benchmark datasets for weakly supervised semantic segmentation.
• The framework is successfully extended to semi-supervised semantic segmentation, yielding state-of-the-art results on two datasets.

Conclusions:

• EPS++ effectively addresses key challenges in weakly supervised semantic segmentation.
• The proposed method offers a robust and efficient approach for improving segmentation accuracy.
• The framework demonstrates versatility and strong performance in both weakly and semi-supervised learning scenarios.