Research map / 3D vision / Depth and stereo
Virtual Scanning: Unsupervised Non-line-of-sight Imaging from Irregularly Undersampled Transients
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Proposes an unsupervised NLOS imaging framework that reconstructs hidden scenes from noisy irregularly undersampled transients using a virtual scanning loss for null-space learning and a SURE-based physics-guided denoiser.
Problem
NLOS imaging from irregularly undersampled transients (IUT) captured over irregular relay surfaces, without paired ground-truth albedo supervision, under noisy low-photon conditions.
Approach
A virtual scanning process re-measures the reconstructed albedo with a different relay-surface mask and enforces consistency between reconstructions, so the network learns null-space components beyond what measurement consistency alone constrains; a Poisson-noise SURE loss with dark counts trains a denoiser without clean data.
Key results
- Ablation: virtual scanning gives an average PSNR gain of 0.89 dB and the SURE denoiser 1.83 dB; full model reaches 20.52 dB PSNR vs 18.69 dB without either component.
- Inference on 128x128x256 IUT: 2.24 s on CPU vs 7.73 h for CC-SOCR, described as about 12,000x faster in CPU mode.
- Qualitative results on simulated and real data show higher fidelity than CC-SOCR, NeTF, USM and direct methods, with the method trained only on alphanumeric simulated data and shutter-like surfaces.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| add the virtual scanning process (ablation) | Ours · Simulated complex-geometry transients, 15 relay surfaces | PSNR: 19.63 -> 20.52 | Helped |
| add the SURE-based denoiser (ablation) | Ours · Simulated complex-geometry transients, 15 relay surfaces | PSNR: 18.69 -> 20.52 | Helped |
| set denoiser ε to 0.1 instead of 1 or 10 (ablation) | SURE-based denoiser · Bunny denoising evaluation | PSNR: 36.97±7.35 -> 41.53 ± 1.25; 34.92±0.17 -> 41.53 ± 1.25 | Helped |
| use VS-loss weight β = 0.001 rather than β = 0 (ablation) | Ours · Bunny reconstruction | PSNR: 16.19±0.21 -> 19.03 ± 0.64 | Helped |
| increase relay-surface interval diversity during training from [4] to [4,8,12,16,20] (ablation) | Ours · Bunny reconstruction | PSNR: 17.63±0.64 -> 19.03 ± 0.64 | Helped |
| increase training rotations from 10 to 40 (ablation) | Ours · Bunny reconstruction | PSNR: 18.08±0.65 -> 19.03 ± 0.64 | Helped |
| increase training rotations from 40 to 60 (ablation) | Ours · Bunny reconstruction | PSNR: 19.03 ± 0.64 -> 18.76±0.75 | Hurt |
| use Ours instead of CC-SOCR | NLOS inference on IUT · IUT, 128 × 128 × 256 (CPU) | runtime: 7.73 h -> 2.24 s | Helped |
Surprising findings
- Applying the SURE denoiser to LCT improved results, and the denoiser's gain (1.83 dB) exceeded that of the virtual scanning component.
Limitations
- Restricted to confocal imaging because no high-speed low-memory non-confocal forward operator is available for deep learning training.
- Only results with networks trained on simulated transients are presented; no training on real transient datasets.
Experimental setup
Datasets: simulated transients from alphanumerics dataset publicly available real NLOS transient datasets [10-12] self-captured NLOS system data
Benchmarks: simulated bunny transients publicly available real NLOS datasets [10-12] self-captured real NLOS data ablation on 1,000 simulated transients with 15 relay surfaces
Metrics: PSNR inference time
Baselines: LCT FK RSD NeTF USM CC-SOCR
Builds on: light-cone transform (LCT) Stein's unbiased risk estimator partial convolution Noise2Noise-style self-supervised denoising
Closest papers by meaning
- Deep Non-line-of-sight Imaging from Under-scanning Measurements (NeurIPS 2023 · similarity 0.91)
- Beyond Volumetric Albedo -- A Surface Optimization Framework for Non-Line-Of-Sight Imaging (CVPR 2019 · similarity 0.88)
- Dual-branch Graph Feature Learning for NLOS Imaging (AAAI 2025 · similarity 0.87)
- WaveMo: Learning Wavefront Modulations to See Through Scattering (CVPR 2024 · similarity 0.85)
- Toward Dynamic Non-Line-of-Sight Imaging with Mamba Enforced Temporal Consistency (NeurIPS 2024 · similarity 0.84)
- Improving Diffusion Posterior Samplers with Lagged Temporal Corrections for Image Restoration (NeurIPS 2026 · similarity 0.83)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.