Dataset for "Neural Networks Meet Light Transport Physics for Passive Non-Line-of-Sight Imaging Enhancement"

This dataset accompanies the paper “Neural Networks Meet Light Transport Physics for Passive Non-Line-of-Sight Imaging Enhancement”, published in IEEE Transactions on Computational Imaging. The dataset was experimentally acquired using an occluder-assisted passive non-line-of-sight (NLOS) imaging system and contains paired NLOS measurements and corresponding ground-truth hidden-scene images. It covers both sparse and complex scenes for the development and evaluation of passive NLOS reconstruction methods. The benchmark datasets include NIST (MNIST and EMNIST), Quickdraw, and SHAPES for sparse-scene reconstruction. The extended datasets include Anime, Supermodel, and STL-10 for complex-scene reconstruction. These datasets were used to evaluate the hybrid physics-data-driven imaging (HPDI) framework proposed in the accompanying paper, including reconstruction fidelity, generalization under distribution shifts, data efficiency, and reconstruction performance across sparse and complex scenes.