Research map / 3D vision / Depth and stereo
Dual-branch Graph Feature Learning for NLOS Imaging
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 DG-NLOS, a dual-branch graph neural network that reconstructs albedo and depth from NLOS measurements in two decoupled training stages, converting dense 3D voxel grids into sparse graph features.
Problem
Non-line-of-sight imaging reconstruction suffers from high computation and memory cost due to 3D grid data, and jointly reconstructing albedo and depth with a mixed loss requires delicate hyperparameter balancing.
Approach
Use a dynamic k-NN graph with negative-vertex graph selection and graph aggregation/update blocks to turn dense NLOS voxel grids into sparse graph features, and decouple albedo and depth reconstruction into an albedo-first branch followed by a depth-focused branch with frozen albedo parameters.
Key results
- On the motorbikes dataset, DG-NLOS reaches 29.93 dB PSNR and 0.92 SSIM, versus 28.77 dB for MIMU and 27.54 dB for LFE.
- DG-NLOS runs in 0.18 s per reconstruction, faster than LFE (0.96 s) and MIMU (1.12 s) in the reported table.
- Albedo-first two-stage training gives 29.93 dB albedo PSNR, above single-branch (28.28 dB) and depth-first (29.65 dB) training.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| use DG-NLOS instead of MIMU | Motorbikes reconstruction · Motorbikes | PSNR (dB): 28.77 -> 29.93 | Helped |
| use DG-NLOS instead of MIMU | Motorbikes reconstruction · Motorbikes | RMSE: 0.06 -> 0.04 | Helped |
| add graph selection to the baseline (ablation) | DG-NLOS baseline framework · Ablation study | PSNR: 28.37 -> 28.68 | Helped |
| add graph update to the baseline (ablation) | DG-NLOS baseline framework · Ablation study | PSNR: 28.37 -> 29.41 | Helped |
| add channel fusion to the baseline (ablation) | DG-NLOS baseline framework · Ablation study | PSNR: 28.37 -> 29.93 | Helped |
| replace ResEdgeConv with GIN or GraphSAGE (ablation) | DG-NLOS graph aggregation · Ablation study | PSNR: 29.93 -> 27.37; 29.93 -> 27.68 | Hurt |
| replace ResEdgeConv with Max-Relative (ablation) | DG-NLOS graph aggregation · Ablation study | PSNR: 29.93 -> 28.42 | Hurt |
| train with a single branch and mixed albedo-depth loss instead of two-stage branches (ablation) | DG-NLOS · Ablation study | Albedo PSNR: 29.93 -> 28.28 | Hurt |
| train the depth branch before the albedo branch (ablation) | DG-NLOS · Ablation study | Albedo PSNR: 29.93 -> 29.65 | Hurt |
| increase the loss scales from 1 to 3 (ablation) | DG-NLOS multi-scale loss · Ablation study | PSNR: 28.41 -> 29.93 | Helped |
| replace L1 loss with MSE or L1+MSE (ablation) | DG-NLOS loss function · Ablation study | PSNR: 29.93 -> 26.90; 29.93 -> 28.63 | Hurt |
Surprising findings
- GIN and GraphSAGE aggregation reduce performance relative to ResEdgeConv in the ablation.
- Its memory use (9.6 GB) is higher than LFE (5.0 GB) and MIMU (5.8 GB), so the claimed memory savings are not uniform across baselines.
Limitations
- Ablations and comparisons are reported only for a few simulated scenes and six real scenes.
- Authors note the model is not yet tested on a wide range of corruptions.
Experimental setup
Datasets: LFE training dataset (3000 simulated measurements) real NLOS measurements (6 scenes)
Benchmarks: motorbikes simulated dataset real NLOS scenes
Metrics: PSNR SSIM RMSE MAD runtime (s) memory (GB) params (M) FLOPs (G)
Baselines: SP FBP LCT FK Phasor LFE MIMU NLOST PnP
Builds on: EdgeConv ResEdgeConv LFE Jump Knowledge Network
Closest papers by meaning
- Virtual Scanning: Unsupervised Non-line-of-sight Imaging from Irregularly Undersampled Transients (NeurIPS 2024 · similarity 0.87)
- LIRM: Large Inverse Rendering Model for Progressive Reconstruction of Shape, Materials and View-dependent Radiance Fields (CVPR 2025 · similarity 0.85)
- GTR: Improving Large 3D Reconstruction Models through Geometry and Texture Refinement (ICLR 2025 · similarity 0.84)
- MS-GS: Multi-Appearance Sparse-View 3D Gaussian Splatting in the Wild (NeurIPS 2025 · similarity 0.84)
- UniSDF: Unifying Neural Representations for High-Fidelity 3D Reconstruction of Complex Scenes with Reflections (NeurIPS 2024 · similarity 0.84)
- ESR-NeRF: Emissive Source Reconstruction Using LDR Multi-view Images (CVPR 2024 · similarity 0.84)
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.