Multi-Task Deep Relative Attribute Learning for Visual Urban Perception.

Area of Science:

• Computer Vision
• Urban Perception
• Machine Learning

Background:

• Visual urban perception quantifies environmental attributes from street-view images.
• Existing methods struggle to fully leverage pairwise comparisons or inter-attribute relationships.
• Current approaches often independently learn perceptual attributes, missing synergistic learning opportunities.

Purpose of the Study:

• To develop a novel network for simultaneous learning of multiple visual urban perceptual attributes.
• To address limitations of existing methods in exploiting pairwise comparisons and attribute interdependencies.
• To enhance urban scene understanding through a unified multi-task learning framework.

Main Methods:

• Proposed a Multi-Task Deep Relative Attribute Learning Network (MTDRALN) using Siamese networks for simultaneous attribute prediction.
• Incorporated structured sparsity to leverage natural attribute grouping based on semantic relatedness.
• Integrated both ranking and classification sub-networks with joint loss functions for discriminative feature learning.

Main Results:

• MTDRALN demonstrated superior performance in learning multiple perceptual attributes concurrently.
• The network effectively utilizes pairwise comparisons and inter-attribute relationships for improved accuracy.
• Experiments on the Place Pulse 2.0 dataset validated the proposed network's effectiveness and ability to learn salient features.

Conclusions:

• The proposed MTDRALN offers an effective approach for multi-attribute visual urban perception.
• Simultaneous learning and structured sparsity significantly enhance the understanding of urban environments.
• The end-to-end trainable network facilitates synergistic deep feature learning and multi-task attribute learning.