Minimal Angular Sampling for Arbitrary-View Light Field Reconstruction Using Spectral-Angular View Selection

Abstract

Dense light field acquisition provides many angular observations of the same scene, but using all available views can introduce redundancy, disparity-related inconsistency, and unnecessary computational cost. This paper addresses the problem of minimal angular sampling for arbitrary-view light field reconstruction. Instead of designing a new reconstruction network, we investigate how to select a compact and informative subset of angular views before reconstruction. The proposed training-free strategy ranks candidate views using spectral similarity, angular proximity, and pretrained feature similarity, and then reconstructs the target view using a fixed fusion-based pipeline. Experiments on benchmark light field scenes show that carefully selected compact view subsets can outperform both target-only reconstruction and full-view fusion. These results demonstrate that reconstruction quality depends more on the relevance of selected angular views than on the total number of available views.