Reusable Architecture Growth for Continual Stereo Matching.

Area of Science:

• Computer Vision
• Machine Learning
• Deep Learning

Background:

• Recent stereo depth estimation models utilize convolutional neural networks for dense disparity regression.
• Continuous acquisition of training data in practical applications necessitates continual learning capabilities for models.
• Existing models often struggle with forgetting previously learned scenes when adapting to new data.

Purpose of the Study:

• To develop a continual stereo matching framework capable of learning new scenes, preventing catastrophic forgetting, and performing continuous disparity prediction.
• To introduce a novel Reusable Architecture Growth (RAG) framework for adaptive and efficient continual learning in stereo matching.
• To enhance the adaptability of stereo depth estimation models for practical, real-world deployment.

Main Methods:

• Proposed a Reusable Architecture Growth (RAG) framework employing task-specific neural unit search and architecture growth.
• Implemented both supervised and self-supervised learning approaches within the RAG framework for continual scene learning.
• Introduced a Scene Router module for adaptive selection of scene-specific architectural paths during inference.

Main Results:

• The RAG framework demonstrated high reusability of previous neural units while achieving strong performance on new scenes.
• The proposed method significantly outperformed state-of-the-art methods, especially in challenging cross-dataset scenarios.
• Experiments confirmed the framework's impressive performance across diverse environmental conditions (weather, road, city).

Conclusions:

• The RAG framework effectively addresses the challenge of continual stereo matching, enabling models to adapt to new scenes without forgetting.
• The Scene Router module enhances inference adaptability, facilitating end-to-end stereo architecture learning.
• This approach holds significant potential for practical deployment in real-world stereo depth estimation applications.