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

Area of Science:

• Computer Vision
• Signal Processing
• Photonics

Background:

• Single-photon sensors offer potential for 3D mapping, autonomous driving, and gesture recognition.
• Low photon counts and Poisson noise significantly challenge high-quality depth reconstruction.

Purpose of the Study:

• To develop a robust network architecture for high-quality scene depth reconstruction from single-photon sensor data.
• To address noise suppression and improve temporal feature robustness under low-photon conditions.

Main Methods:

• Proposed a network integrating dense encoding and nested decoding.
• Introduced Asymmetric Gaussian Convolution (AGC) for temporal feature enhancement and Sparse Channel Dense Block (SCDB) for feature extraction.
• Developed a Temporal Similarity Aggregation (TSA) module for improved spatial feature correlation using photon temporal distribution.

Main Results:

• The proposed network effectively suppresses noise and achieves high-quality depth reconstruction.
• Demonstrated significant performance improvements over existing methods on synthetic datasets.
• Showcased robust generalization to real-world imaging systems under low photon flux.

Conclusions:

• The novel network architecture significantly enhances depth reconstruction accuracy for single-photon sensors.
• The developed methods provide a robust solution for low-photon imaging challenges.
• This work advances 3D mapping and perception in low-light environments.