Research map / 3D vision / Depth and stereo
A Generative Model for Generic Light Field Reconstruction
Paramanand Chandramouli, Kanchana Vaishnavi Gandikota, Andreas Görlitz, Andreas Kolb, Michael Moeller
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 |
|---|---|---|---|
| average over overlapping patches with stride 5 (ablation) | CVAE prior · 3x3 to 7x7 view synthesis | PSNR (dB): 38.53 -> 39.77 | Improved |
| average over overlapping patches with stride 5 (ablation) | CVAE prior · 5x5 view synthesis, mask M1, Dino | PSNR (dB): 39.57 -> 41.53 | Improved |
| average over overlapping patches with stride 5 (ablation) | CVAE prior · 5x5 coded aperture, mask M1, Dino | PSNR (dB): 34.97 -> 38.46 | Improved |
| use CVAE prior with energy minimization instead of dictionary prior | energy-minimization LF reconstruction · 5x5 view synthesis, mask M1 | PSNR (dB): by 2.63 dB (mask M1), 3.13 dB (mask M2) | Improved |
| use CVAE prior with overlapping patches instead of dictionary prior | energy-minimization LF reconstruction · 5x5 view synthesis | PSNR (dB): by 4 dB and 4.4 dB | Improved |
| reduce known views from 3x3 to 5 randomly chosen views (ablation) | CVAE prior with overlapping patches · 7x7 view synthesis | PSNR (dB): 39.77 -> 39.57 | Hurt |
| use CVAE prior instead of end-to-end CNN with noisy inputs (sigma 0.1) | CVAE prior · Cars 3x3 to 7x7, Gaussian noise 0.1 | PSNR (dB): 29.95 vs 31.67 | Improved |
| use CVAE prior instead of end-to-end CNN with 50% pixel drop | CVAE prior · Cars 3x3 to 7x7, 50% pixel drop | PSNR (dB): 13.60 vs 31.68 | Improved |
| use L1 loss instead of MSE loss for salt and pepper noise (ablation) | CVAE prior · Cars, salt and pepper noise | PSNR (dB): PSNR dropped by about 2 dB with MSE vs 0.1 dB with L1 | Improved |
| alter coded mask with swaps at test time (1 swap) | end-to-end network · Dino coded aperture M1 | PSNR (dB): 38.7 -> 24.3 | Hurt |
| alter coded mask with swaps at test time (3 swaps) | CVAE prior with overlapping patches · Dino coded aperture M1 | PSNR (dB): 38.46 -> 38.19 | No clear effect |
| add TV penalty on central view when all views are corrupted | CVAE prior · Cars, all views corrupted | PSNR (dB) | Improved |
Closest papers by meaning
- Light Field Neural Rendering (CVPR 2022 · similarity 0.90)
- High-Dimensional Dense Residual Convolutional Neural Network for Light Field Reconstruction (TPAMI 2019 · similarity 0.90)
- Learning Light Field Angular Super-Resolution via a Geometry-Aware Network (AAAI 2020 · similarity 0.89)
- Light Field Super-Resolution using a Low-Rank Prior and Deep Convolutional Neural Networks (TPAMI 2019 · similarity 0.89)
- Deep Spatial-angular Regularization for Compressive Light Field Reconstruction over Coded Apertures (ECCV 2020 · similarity 0.89)
- NeLF-Pro: Neural Light Field Probes for Multi-Scale Novel View Synthesis (CVPR 2024 · 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.
