CCNet: Criss-Cross Attention for Semantic Segmentation.

Area of Science:

• Computer Vision
• Deep Learning
• Artificial Intelligence

Background:

• Contextual information is crucial for accurate visual understanding in tasks like semantic segmentation and object detection.
• Existing methods for capturing context can be computationally expensive and memory-intensive.

Purpose of the Study:

• To propose an effective and efficient Criss-Cross Network (CCNet) for obtaining full-image contextual information.
• To improve the performance of visual understanding tasks by better leveraging contextual dependencies.

Main Methods:

• Introduced a novel criss-cross attention module that allows each pixel to gather information from its criss-cross path.
• Employed a recurrent operation to enable pixels to capture full-image dependencies.
• Proposed a category consistent loss to enhance feature discriminability.

Main Results:

• CCNet demonstrates significant improvements in GPU memory efficiency (11x less than non-local blocks) and computational efficiency (85% reduction in FLOPs).
• Achieved state-of-the-art performance on multiple benchmarks: Cityscapes (81.9% mIoU), ADE20K (45.76% mIoU), and LIP (55.47% mIoU).
• Code is publicly available for reproducibility and further research.

Conclusions:

• CCNet offers a highly effective and efficient solution for incorporating full-image context into visual understanding models.
• The proposed method sets new state-of-the-art benchmarks in semantic segmentation, human parsing, and instance segmentation.
• CCNet's efficiency and performance make it a valuable contribution to the field of computer vision.