Convolutional Oriented Boundaries: From Image Segmentation to High-Level Tasks.

Area of Science:

• Computer Vision
• Machine Learning
• Image Analysis

Background:

• Generic image classification Convolutional Neural Networks (CNNs) are prevalent but often lack detailed contour and hierarchical region information.
• Existing methods for contour detection and hierarchical segmentation can be computationally intensive and may not generalize well.

Purpose of the Study:

• To introduce Convolutional Oriented Boundaries (COB), a novel method for generating multiscale oriented contours and region hierarchies.
• To demonstrate the computational efficiency and superior performance of COB compared to state-of-the-art methods.
• To validate the generalization capabilities of COB across diverse datasets and unseen categories.

Main Methods:

• Leveraging generic image classification CNNs as a foundation for contour and hierarchy generation.
• Implementing a single CNN forward pass for efficient multi-scale contour detection.
• Utilizing a novel sparse boundary representation for hierarchical segmentation.

Main Results:

• COB achieves state-of-the-art performance in boundary detection and region hierarchy generation across multiple datasets (BSDS, PASCAL Context, PASCAL Segmentation, NYUD).
• Estimating both contour strength and orientation significantly improves accuracy.
• COB enhances performance in high-level computer vision tasks, including object proposals, semantic contours, semantic segmentation, and object detection (MS-COCO, SBD, PASCAL).

Conclusions:

• COB offers a computationally efficient and highly effective approach for extracting detailed structural information from images.
• The method demonstrates strong generalization, making it applicable to a wide range of computer vision tasks.
• COB represents a significant advancement in image analysis, providing accurate contours and hierarchical segmentations.