Multi-scale photon-efficient depth reconstruction based on spatiotemporal priors

Scene depth reconstruction using single-photon sensors has broad applications in 3D mapping, autonomous driving, and gesture recognition. However, high-quality depth reconstruction is always challenged by low photon counts and poor signal-to-noise ratios, particularly due to the Poisson noise inherent in photon-starved detection. In this paper, we propose a network architecture that integrates dense encoding and nested decoding to suppress noise and achieve high-quality depth reconstruction. For the encoding part, we design an Asymmetric Gaussian Convolution (AGC) to enhance temporal features and mitigate redundancy, while a Sparse Channel Dense Block (SCDB) is employed for the extraction of effective features. Regarding the decoding part, cross-layer connections are adopted to enable the reuse of multi-scale spatiotemporal features. Additionally, we introduce a Temporal Similarity Aggregation (TSA) module, which takes the temporal distribution of photons as the basis for spatial feature correlation, thereby improving the robustness of temporal features under low-photon conditions. Extensive experiments demonstrate that the proposed 3D reconstruction network significantly outperforms existing methods on synthetic datasets and exhibits robust generalization to real-world imaging systems under low photon flux.