Research map / 3D vision / Depth and stereo
V2V3D: View-to-View Denoised 3D Reconstruction for Light Field Microscopy
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 V2V3D, an unsupervised view-to-view framework that jointly denoises and reconstructs 3D volumes from light field microscopy images using split-view noise2noise supervision and wave-optics PSF-based feature alignment.
Problem
Light field microscopy 3D reconstruction is highly sensitive to sensor noise, and existing learning-based methods need hard-to-obtain ground-truth annotated data; pixel-matched noise2noise denoising is inapplicable because views are coordinate-unmatched.
Approach
Split LF views into two disjoint subsets, reconstruct a volume from each with a U-Net, and use forward PSF projection to simulate the other subset's views as noise-independent supervision; additionally convert PSFs into alignment kernels for back-projecting features into a shared 3D space, plus an FFT loss and a de-crosstalk loss.
Key results
- On the synthetic dataset V2V3D reaches 39.05 PSNR / 0.898 SSIM average, versus 37.71 / 0.792 for VCDNet and 32.19 / 0.721 for RLD.
- Ablation: removing the view2view framework drops PSNR from 39.05 to 30.81; removing wave-optics alignment drops it to 38.25.
- Runtime 0.413 s per volume for V2V3D versus 7.34 s for RLD and 61.4 s for DINER, though with 210.77M parameters versus 87.98M for VCDNet.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| use V2V3D instead of VCDNet | LFM reconstruction · Synthetic dataset (Average) | PSNR: 37.71 -> 39.05 | Helped |
| use V2V3D instead of VCDNet | LFM reconstruction · Synthetic dataset (Average) | SSIM: 0.792 -> 0.898 | Helped |
| add the view2view framework (ablation) | V2V3D · Synthetic dataset (Average) | PSNR: 30.81 -> 39.05 | Helped |
| add feature alignment (ablation) | V2V3D · Synthetic dataset (Average) | PSNR: 38.25 -> 39.05 | Helped |
| add FFT loss (ablation) | V2V3D · Synthetic dataset (Average) | PSNR: 37.23 -> 39.05 | Helped |
| add de-crosstalk loss (ablation) | V2V3D · Synthetic dataset (Average) | PSNR: 36.09 -> 39.05 | Helped |
Surprising findings
- V2V3D is slower than VCDNet (0.413 s vs 0.047 s) and has more parameters (210.77M vs 87.98M), though faster than RLD and DINER.
- Applying DeepCAD pre-denoising before RLD or DINER can introduce more artifacts, due to signal crosstalk from adjacent-view N2N pairing.
Limitations
- Relies on an accurate imaging system (PSF) model; PSF is not optimized during training.
- Uses simple averaging to fuse the two branch volumes; more advanced fusion was not found to help.
- Real-world evaluation is qualitative since no ground truth is available.
Experimental setup
Datasets: synthetic LFM dataset built from 2pSAM PSFs and fluorescence volumes real-world 2pSAM LFI recordings (brain slice, neutrophils) LLFF Fern scene
Benchmarks: V2V3D LFM synthetic dataset (11 test scenes) real-world LFM brain slice and neutrophil recordings LLFF Fern macro scenario
Metrics: PSNR SSIM runtime (s) parameters (M)
Baselines: RLD VCDNet DINER DeepCAD+RLD DeepCAD+DINER
Builds on: noise2noise Richardson-Lucy deconvolution DINER VCDNet DeepCAD
Closest papers by meaning
- From Pixels to Views: Learning Angular-Aware and Physics-Consistent Representations for Light Field Microscopy (NeurIPS 2025 · similarity 0.90)
- Volume Tells: Dual Cycle-Consistent Diffusion for 3D Fluorescence Microscopy De-noising and Super-Resolution (CVPR 2025 · similarity 0.85)
- OpticalNet: An Optical Imaging Dataset and Benchmark Beyond the Diffraction Limit (CVPR 2025 · similarity 0.85)
- MicroDiffusion: Implicit Representation-Guided Diffusion for 3D Reconstruction from Limited 2D Microscopy Projections (CVPR 2024 · similarity 0.84)
- A Gaussian Filter-Based 3D Registration Method for Series Section Electron Microscopy (AAAI 2025 · similarity 0.84)
- S2D-LFE: Sparse-to-Dense Light Field Event Generation (CVPR 2025 · 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.