Research map / 3D vision / Depth and stereo
Inertial Algorithm with Dry Fraction and Convolutional Sparse Coding for 3D Localization with 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 Fast-IPGDF, an inertial proximal gradient algorithm with dry friction and Nesterov-style acceleration for slice-based convolutional sparse coding, to improve EPI-based 3D localization of fluorescent beads in light field microscopy.
Problem
Neuron/point-source 3D localization from single light field microscopy images is less accurate at greater depths and in noisy conditions, and ADMM-based convolutional sparse coding solvers can converge to local minima and are slow.
Approach
Replaces ADMM in the convolutional sparse coding model with an accelerated inertial proximal gradient method using a dry friction term (Heavy Ball with Dry Friction), with a slice-based CSC formulation, to avoid local minima and speed convergence for depth matching on EPIs.
Key results
- Fast-IPGDF processes a single EPI in 2.235 s versus 6.489 s for the ADMM-based method, about 2.9x faster.
- Under Poisson noise (PSNR ~24 dB), the ADMM-based method's localization curve fluctuates and fails, while Fast-IPGDF localizes depth accurately with minimal x/y error.
- Best RMSE of 2.11e-6 at dry friction coefficient r=0.05; RMSE rises to 3.97e-6 at viscous damping γ=55.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| replace the ADMM-based method with Fast-IPGDF | CSC-based EPI localization · Single EPI | time cost: 6.489 s -> 2.235 s | Helped |
| set dry friction coefficient r to 0.01 instead of 0.05 (ablation) | Fast-IPGDF · Localization experiment | RMSE (10−6): 2.11 -> 2.41 | Hurt |
| set dry friction coefficient r to 0.03 instead of 0.05 (ablation) | Fast-IPGDF · Localization experiment | RMSE (10−6): 2.11 -> 2.35 | Hurt |
| set dry friction coefficient r to 0.07 instead of 0.05 (ablation) | Fast-IPGDF · Localization experiment | RMSE (10−6): 2.11 -> 2.76 | Hurt |
| set viscous damping parameter γ to 10 instead of 30 (ablation) | Fast-IPGDF · Localization experiment | RMSE (10−6): 2.41 -> 2.59 | Hurt |
| increase viscous damping parameter γ from 30 to 55 (ablation) | Fast-IPGDF · Localization experiment | RMSE (10−6): 2.41 -> 3.97 | Hurt |
| increase viscous damping parameter γ from 30 to 65 (ablation) | Fast-IPGDF · Localization experiment | RMSE (10−6): 2.41 -> 5.83 | Hurt |
Surprising findings
- Increasing the viscous damping parameter γ beyond a point degrades RMSE (5.83e-6 at γ=65), and larger γ also increases iteration count.
Limitations
- Accuracy decreases as depth increases, and experiments use simulated/non-scattered fluorescent beads only.
- Convergence results are derived from prior theory and evaluated on a limited set of synthetic bead experiments.
Experimental setup
Datasets: simulated fluorescent bead light field images (10 µm beads in agarose)
Benchmarks: fluorescent bead localization (depth -55 to +55 µm) noisy-scene localization (Poisson noise, PSNR ~24 dB) dual light source localization
Metrics: RMSE localization error running time per EPI sparsity of codes
Baselines: ADMM-based CSC method (Song et al. 2020) CSC method (Song et al. 2021)
Builds on: Heavy Ball with Dry Friction dynamical system (Adly and Attouch 2022) convolutional sparse coding LFM model (Song et al. 2021) ADMM
Closest papers by meaning
- From Pixels to Views: Learning Angular-Aware and Physics-Consistent Representations for Light Field Microscopy (NeurIPS 2025 · similarity 0.82)
- V2V3D: View-to-View Denoised 3D Reconstruction for Light Field Microscopy (CVPR 2025 · similarity 0.82)
- MicroDiffusion: Implicit Representation-Guided Diffusion for 3D Reconstruction from Limited 2D Microscopy Projections (CVPR 2024 · similarity 0.81)
- A Gaussian Filter-Based 3D Registration Method for Series Section Electron Microscopy (AAAI 2025 · similarity 0.81)
- CodedEvents: Optimal Point-Spread-Function Engineering for 3D-Tracking with Event Cameras (CVPR 2024 · similarity 0.80)
- Nonlinear dynamics of localization in neural receptive fields (NeurIPS 2024 · similarity 0.79)
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.