Research map / 3D vision / Depth and stereo
PolarRec: Improving Radio Interferometric Data Reconstruction Using Polar Coordinates
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 PolarRec, a transformer-conditioned neural-field pipeline that encodes radio interferometric visibility samples in polar coordinates, using a radial-weighted loss and angular-group tokens to reconstruct all frequency components more accurately and efficiently.
Problem
Reconstructing sparse radio interferometric visibility samples into dense form before imaging: existing methods miss high-frequency components, producing blurred edges and artifacts, and Transformer encoding of every point is computationally expensive due to data skew.
Approach
Convert visibility samples to polar coordinates, weight a visibility loss by radial distance (Radial Visibility Loss) to emphasize high frequencies, and group points by polar angle into group tokens before Transformer encoding, conditioning a neural field via FiLM.
Key results
- PolarRec achieves PSNR of 24.21 (MG), 26.54 (IRSG), 25.43 (UTSG), 25.88 (EGB) versus 20.56, 24.10, 21.32, 21.28 for Neural Interferometry, with lower LFD on all four datasets.
- At batch size 32 with 1660 points, PolarRec latency is 5.37 ms versus 232.04 ms for Neural Interferometry (GFLOPs 24.40 vs 352.16 at group size 32).
- Angular-coordinate grouping outperforms clustering and radial grouping across group sizes; removing either RVL weight term degrades all metrics.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| use PolarRec instead of Neural Interferometry | radio visibility reconstruction · MG, IRSG, UTSG, EGB | PSNR: MG: 20.560 (2.287) -> 24.206 (2.547); IRSG: 24.099 (2.760) -> 26.540 (2.653); UTSG: 21.323 (2.303) -> 25.430 (2.565); EG | Helped |
| use PolarRec instead of Neural Interferometry | radio visibility reconstruction · MG, IRSG, UTSG, EGB | LFD: MG: 1.229 (0.282) -> 0.924 (0.284); IRSG: 0.916 (0.302) -> 0.666 (0.238); UTSG: 1.038 (0.313) -> 0.707 (0.241); EGB: 1.1 | Helped |
| add radial weight w2 to RVL (ablation) | PolarRec · MG, IRSG, UTSG, EGB | PSNR: MG: 22.498 (2.578) -> 24.206 (2.547); IRSG: 24.435 (2.577) -> 26.540 (2.653); UTSG: 23.477 (2.574) -> 25.430 (2.565); EG | Helped |
| add focal weight w1 to RVL (ablation) | PolarRec · MG, IRSG, UTSG, EGB | PSNR: MG: 22.706 (2.797) -> 24.206 (2.547); IRSG: 25.213 (2.641) -> 26.540 (2.653); UTSG: 24.393 (2.627) -> 25.430 (2.565); EG | Helped |
| increase group size from 32 to 128 (ablation) | PolarRec · MG, IRSG, UTSG, EGB | PSNR: MG: 23.51 -> 22.06; IRSG: 26.80 -> 26.02; UTSG: 25.63 -> 24.48; EGB: 26.07 -> 25.30 | Mixed |
| increase group size from 1 to 32 (ablation) | PolarRec · MG, IRSG, UTSG, EGB | GFLOPs: 366.14 -> 24.40 | Helped |
| group visibility points by angular coordinate rather than clustering or radial coordinate (ablation) | PolarRec · MG, IRSG, UTSG, EGB | LFD | Helped |
Surprising findings
- Increasing group size beyond 32 slightly reduces PSNR and SSIM, indicating a quality-efficiency trade-off.
Limitations
- Generalizability to other data is limited; reconstructed real-data images still contain some artifacts.
- Evaluated only on simulated galaxy-image-derived visibility data with a fixed 8-telescope EHT configuration.
- Group size above 32 degrades reconstruction quality slightly.
Experimental setup
Datasets: Galaxy10 DECaLS (Merging Galaxies, In-between Round Smooth Galaxies, Unbarred Tight Spiral Galaxies, Edge-on Galaxies wi eht-imaging simulated EHT 8-telescope visibilities real observational data [9]
Benchmarks: Galaxy10 DECaLS four-morphology visibility reconstruction
Metrics: LFD PSNR SSIM latency (ms) GFLOPs
Baselines: CLEAN U-Net Radionets Neural Interferometry
Builds on: Neural Interferometry Focal Frequency Loss conditional neural fields FiLM
Closest papers by meaning
- VisRec: A Semi-Supervised Approach to Visibility Data Reconstruction in Radio Astronomy (AAAI 2025 · similarity 0.88)
- PolarScale: A Physics-Grounded Benchmark for Radiometrically Consistent RGB-to-Stokes Estimation (NeurIPS 2026 · similarity 0.80)
- PolarFree: Polarization-based Reflection-Free Imaging (CVPR 2025 · similarity 0.80)
- GNeRP: Gaussian-guided Neural Reconstruction of Reflective Objects with Noisy Polarization Priors (ICLR 2024 · similarity 0.80)
- Polarization Wavefront Lidar: Learning Large Scene Reconstruction from Polarized Wavefronts (CVPR 2024 · similarity 0.80)
- High-Fidelity Polarimetric Implicit 3D Reconstruction with View-Dependent Physical Representation (AAAI 2025 · 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.