Research map / Architectures and efficiency / RNNs and normalization
Learnable Burst-Encodable Time-of-Flight Imaging for High-Fidelity Long-Distance Depth Sensing
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 Burst-Encodable ToF (BE-ToF), a burst-mode single-frequency iToF paradigm with an end-to-end learned binarized coding function and a Restormer-based network (RSCF-Net), validated on simulated NYU-V2 and a outdoor prototype.
Problem
Indirect ToF imaging suffers from phase wrapping and low SNR at long distances, while direct ToF needs ultra-short pulses and high-resolution TDCs; existing fixes need multi-frequency modulation or scene priors.
Approach
Emit burst-mode light pulses so the phase shift spans the full 0-2π within one long burst period, avoiding phase wrapping with single-frequency modulation; jointly learn binarized coding functions (constrained by a double-well loss and first-order difference loss) and a Restormer-based depth reconstruction network, with a Fisher-information guidance loss.
Key results
- At 90-93m under high SNR, the method reaches MAE 18.96 mm vs 75.19 mm for FisherToF and 117.21 mm for Sine+PS.
- At 60-63m under low SNR, MAE 26.08 mm vs 54.73 mm for FisherToF.
- Replacing the network with RSCF-Net while keeping learned coding fixed lowers MAE relative to DeepToF, MaskToF and FisherToF in the ablation tables.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| use BE-ToF instead of the strongest listed baseline | Depth sensing method · 0-3m | MAE (mm): 7.19 -> 5.90 (H_snr); 10.46 -> 6.95 (M_snr); 16.91 -> 12.71 (L_snr) | Helped |
| use BE-ToF instead of the strongest listed baseline | Depth sensing method · 30-33m | MAE (mm): 20.80 -> 8.03 (H_snr); 24.54 -> 12.25 (M_snr); 31.61 -> 18.29 (L_snr) | Helped |
| use BE-ToF instead of the strongest listed baseline | Depth sensing method · 60-63m | MAE (mm): 31.50 -> 11.93 (H_snr); 35.02 -> 16.60 (M_snr); 45.54 -> 26.08 (L_snr) | Helped |
| use BE-ToF instead of the strongest listed baseline | Depth sensing method · 90-93m | MAE (mm): 42.64 -> 18.96 (H_snr); 45.12 -> 21.99 (M_snr); 56.97 -> 29.58 (L_snr) | Helped |
| use learnable coding functions instead of square coding (ablation) | BE-ToF with RSCF-Net · 0-3m, H_snr | MAE (mm): 12.66 -> 5.90 | Helped |
| use RSCF-Net instead of DeepToF (ablation) | BE-ToF with learned coding functions · 0-3m, H_snr | MAE (mm): 14.76 -> 5.90 | Helped |
| add Fisher guidance loss (ablation) | BE-ToF learnable framework · Synthetic assessment, different distances and SNRs | MAE (mm) | Helped |
| add double well loss (ablation) | BE-ToF learnable coding functions · Coding-function hardware implementability | hardware implementability | Helped |
| add first-order difference loss (ablation) | BE-ToF learnable coding functions · Coding-function hardware implementability | hardware implementability | Helped |
| add CFEB (ablation) | RSCF-Net · Synthetic assessment, different distances and SNRs | MAE (mm) | Helped |
| add MFFB (ablation) | RSCF-Net · Synthetic assessment, different distances and SNRs | MAE (mm) | Helped |
Limitations
- Imaging range is confined to a window between c·τ/2 and c·(τ+T_m)/2, and higher precision narrows this range.
- Evaluation on simulated NYU-V2 data with synthetic physical model; real-world validation is a single outdoor scene.
- Learned coding functions are constrained to binary states for hardware implementability.
Experimental setup
Datasets: NYU-V2
Benchmarks: NYU-V2 simulated ToF at 0-3m, 30-33m, 60-63m, 90-93m
Metrics: MAE (mm)
Baselines: Sine+PS Square+PS DeepToF FisherToF MaskToF
Builds on: Restormer FisherToF DeepToF MaskToF Efficient Channel Attention
Closest papers by meaning
- Deep End-to-End Time-of-Flight Imaging (CVPR 2018 · similarity 0.89)
- Snapshot Lidar: Fourier Embedding of Amplitude and Phase for Single-Image Depth Reconstruction (CVPR 2024 · similarity 0.87)
- Separating the Wheat from the Chaff: Spatio-Temporal Transformer with View-interweaved Attention for Photon-Efficient Depth Sensing (AAAI 2025 · similarity 0.85)
- SVDC: Consistent Direct Time-of-Flight Video Depth Completion with Frequency Selective Fusion (CVPR 2025 · similarity 0.84)
- iToF-flow-based High Frame Rate Depth Imaging (CVPR 2024 · similarity 0.83)
- Time-Efficient Light-Field Acquisition Using Coded Aperture and Events (CVPR 2024 · 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.