Research map / 3D vision / Depth and stereo
Physics to the Rescue: Deep Non-Line-of-Sight Reconstruction for High-Speed Imaging
Fangzhou Mu, Sicheng Mo, Jiayong Peng, Xiaochun Liu, Ji Hyun Nam, Siddeshwar Raghavan, Andreas U. Velten, Yin Li
Read the paperCheck your idea against 137,217 papers
AI-extracted. The measured effects below were read by AI from the paper's full text (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| add supervised physics-embedded model instead of LFE | Ours (sup) · Alphanumerics unseen poses | PSNR: 22.42 -> 24.60 | Improved |
| add unsupervised training (Poisson loss on transients) to supervised training (ablation) | Ours (joint) · Alphanumerics unseen poses | PSNR: 24.60 -> 25.05 | Improved |
| add supervised loss to unsupervised training (ablation) | Ours (unsup) · Alphanumerics unseen poses | SSIM: 0.709 -> 0.861 | Improved |
| use unsupervised transient training only vs RSD | Ours (unsup) · Alphanumerics unseen poses | PSNR: 20.83 -> 20.97 | No clear effect |
| joint training instead of supervised only (ablation) | Ours (joint) · CMU out-of-distribution | PSNR: 22.46 -> 22.11 | Hurt |
| unsupervised-only training instead of supervised for out-of-distribution (ablation) | Ours (unsup) · CMU out-of-distribution | PSNR: 22.46 -> 23.18 | Improved |
| replace NeTF rendering with proposed transient rendering recipe (NeTF++) | NeTF · CMU out-of-distribution | PSNR: 20.31 -> 23.26 | Improved |
| feed-forward learned scene priors instead of per-scene NeTF++ | Ours (sup) · CMU out-of-distribution | PSNR: 23.26 -> 22.46 | Hurt |
| use ConvNet-free volume renderer instead of LFE on motorbikes | Ours (sup) · Motorbikes unseen poses | PSNR: 21.09 -> 24.05 | Improved |
| use our encoder features instead of RSD features for classification | Ours (sup) · Alphanumerics classification | accuracy: 82.2% -> 85.4% | Improved |
| add joint training for classification features (ablation) | Ours (joint) · Alphanumerics classification | accuracy: 85.4% -> 85.2% | No clear effect |
| add total variation prior (ablation) | Ours · NLOS intensity rendering | qualitative | Improved |
Closest papers by meaning
- Enhancing Non-line-of-sight Imaging via Learnable Inverse Kernel and Attention Mechanisms (ICCV 2023 · similarity 0.91)
- Deep Non-line-of-sight Imaging from Under-scanning Measurements (NeurIPS 2023 · similarity 0.91)
- Non-Line-of-Sight Imaging With Signal Superresolution Network (CVPR 2023 · similarity 0.90)
- Acoustic Non-Line-Of-Sight Imaging (CVPR 2019 · similarity 0.89)
- Virtual Scanning: Unsupervised Non-line-of-sight Imaging from Irregularly Undersampled Transients (NeurIPS 2024 · similarity 0.89)
- Dual-branch Graph Feature Learning for NLOS Imaging (AAAI 2025 · similarity 0.87)
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.
