REDARTS: Regressive Differentiable Neural Architecture Search for Exploring Optimal Light Field Disparity Estimation Network

Abstract

Deep learning is widely used in various fields of computer vision applications. However, the majority of these state-of-the-art deep learning architectures are computationally expensive and hand-engineered, requiring substantial expertise to discover. Recently, neural architecture search has gained significant attention as an automated tool for constructing deep neural networks. Although it has found optimal architecture for various applications, their impact on light field disparity estimation is just to be investigated. This paper introduces the Regressive Differentiable Neural Architecture Search algorithm, which finds the optimal architecture by preserving search and evaluation architecture dimensions and proposes an adaptive dynamic drop strategy based on candidate operation stability for optimization. Furthermore, the parameter-sharing technique facilitates search super-network with rapid convergence to assist in better systematic decision-making, enhancing the overall efficiency of the algorithm. The evaluation is conducted on two diverse computer vision tasks to demonstrate the generalizability of the proposed search strategy. The proposed search strategy discovers an optimal architecture 2.46 times faster, which achieves performance comparable to recent state-of-the-art deep learning architectures. By reducing the time and effort required to find sub-optimal architecture, this study opens up new opportunities for the research community. It could make advanced computer vision more accessible in complex applications, including light field technology.