DECNet: Dense embedding contrast for unsupervised semantic segmentation.

Area of Science:

• Computer Vision
• Machine Learning
• Artificial Intelligence

Background:

• Unsupervised semantic segmentation aims to classify image pixels without manual annotation.
• Vision transformers pre-trained via self-supervision show promise but often require complex methods.
• Existing approaches can be overly complicated, necessitating simpler alternatives.

Purpose of the Study:

• To develop a straightforward and effective method for unsupervised semantic segmentation.
• To introduce a novel network, Dense Embedding Contrast network (DECNet), for this task.
• To enhance dense representation learning through contrastive methods.

Main Methods:

• Proposing a simple Dense Embedding Contrast network (DECNet).
• Introducing a Nearest Neighbor Similarity (NNS) strategy for creating effective positive and negative pairs in dense contrastive learning.
• Optimizing the Ortho-InfoNCE contrastive objective to mitigate the false negative problem.

Main Results:

• DECNet demonstrates superior performance on unsupervised semantic segmentation tasks.
• The proposed NNS strategy effectively establishes well-defined pairs for contrastive learning.
• Ortho-InfoNCE significantly enhances dense representations by addressing false negatives.
• Experiments on COCO-Stuff and Cityscapes datasets confirm state-of-the-art results.

Conclusions:

• DECNet offers a simple yet powerful approach to unsupervised semantic segmentation.
• The combination of NNS and Ortho-InfoNCE provides a robust framework for dense contrastive learning.
• This work advances the field by providing an efficient and effective solution without complex architectures.