Panorama

Panorama Multi-Focus Fusion

Join the Competition Contribute to Benchmark

Standard reconstruction benchmark — forward model perfectly known, no calibration needed. Score = 0.5 × clip((PSNR−15)/30, 0, 1) + 0.5 × SSIM

# Method Score PSNR (dB) SSIM Source
🥇 PanoFormer 0.808 35.0 0.950 ✓ Certified Image stitching transformer, 2024
🥈 UDIS 0.760 33.0 0.920 ✓ Certified Nie et al., ICCV 2021
🥉 APAP 0.667 29.5 0.850 ✓ Certified Zaragoza et al., CVPR 2013
4 SIFT-RANSAC 0.553 26.0 0.740 ✓ Certified Lowe, IJCV 2004

Dataset: PWM Benchmark (4 algorithms)

Blind Reconstruction Challenge — forward model has unknown mismatch, must calibrate from data. Score = 0.4 × PSNR_norm + 0.4 × SSIM + 0.2 × (1 − ‖y − Ĥx̂‖/‖y‖)

# Method Overall Score Public
PSNR / SSIM
 Dev
PSNR / SSIM
 Hidden
PSNR / SSIM
 Trust Source
🥇 PanoFormer + gradient 0.727
0.784
32.79 dB / 0.951
0.726
29.83 dB / 0.914
0.671
26.89 dB / 0.855
✓ Certified Image stitching transformer, 2024
🥈 UDIS + gradient 0.666
0.755
30.81 dB / 0.928
0.649
25.01 dB / 0.803
0.594
23.86 dB / 0.764
✓ Certified Nie et al., ICCV 2021
🥉 APAP + gradient 0.656
0.699
27.71 dB / 0.875
0.655
25.26 dB / 0.810
0.614
23.71 dB / 0.758
✓ Certified Zaragoza et al., CVPR 2013
4 SIFT-RANSAC + gradient 0.593
0.625
24.21 dB / 0.776
0.599
23.01 dB / 0.731
0.556
21.76 dB / 0.680
✓ Certified Lowe, IJCV 2004

Complete score requires all 3 tiers (Public + Dev + Hidden).

Join the competition →
Scoring: 0.4 × PSNR_norm + 0.4 × SSIM + 0.2 × (1 − ‖y − Ĥx̂‖/‖y‖) PSNR 40% · SSIM 40% · Consistency 20%
Public 5 scenes

Full-access development tier with all data visible.

What you get & how to use

What you get: Measurements (y), ideal forward operator (H), spec ranges, ground truth (x_true), and true mismatch spec.

How to use: Load HDF5 → compare reconstruction vs x_true → check consistency → iterate.

What to submit: Reconstructed signals (x_hat) and corrected spec as HDF5.

Public Leaderboard
# Method Score PSNR SSIM
1 PanoFormer + gradient 0.784 32.79 0.951
2 UDIS + gradient 0.755 30.81 0.928
3 APAP + gradient 0.699 27.71 0.875
4 SIFT-RANSAC + gradient 0.625 24.21 0.776
Spec Ranges (3 parameters)
Parameter Min Max Unit
focus_step -2.0 4.0 μm
registration -0.5 1.0 pixels
exposure_variation -3.0 6.0 %
View Details →
Dev 5 scenes

Blind evaluation tier — no ground truth available.

What you get & how to use

What you get: Measurements (y), ideal forward operator (H), and spec ranges only.

How to use: Apply your pipeline from the Public tier. Use consistency as self-check.

What to submit: Reconstructed signals and corrected spec. Scored server-side.

Dev Leaderboard
# Method Score PSNR SSIM
1 PanoFormer + gradient 0.726 29.83 0.914
2 APAP + gradient 0.655 25.26 0.81
3 UDIS + gradient 0.649 25.01 0.803
4 SIFT-RANSAC + gradient 0.599 23.01 0.731
Spec Ranges (3 parameters)
Parameter Min Max Unit
focus_step -2.4 3.6 μm
registration -0.6 0.9 pixels
exposure_variation -3.6 5.4 %
Submit Reconstruction View Details →
Hidden 5 scenes

Fully blind server-side evaluation — no data download.

What you get & how to use

What you get: No data downloadable. Algorithm runs server-side on hidden measurements.

How to use: Package algorithm as Docker container / Python script. Submit via link.

What to submit: Containerized algorithm accepting y + H, outputting x_hat + corrected spec.

Hidden Leaderboard
# Method Score PSNR SSIM
1 PanoFormer + gradient 0.671 26.89 0.855
2 APAP + gradient 0.614 23.71 0.758
3 UDIS + gradient 0.594 23.86 0.764
4 SIFT-RANSAC + gradient 0.556 21.76 0.68
Spec Ranges (3 parameters)
Parameter Min Max Unit
focus_step -1.4 4.6 μm
registration -0.35 1.15 pixels
exposure_variation -2.1 6.9 %
Submit Algorithm View Details →

Blind Reconstruction Challenge

Challenge

Given measurements with unknown mismatch and spec ranges (not exact params), reconstruct the original signal. A method must be evaluated on all three tiers for a complete score. Scored on a composite metric: 0.4 × PSNR_norm + 0.4 × SSIM + 0.2 × (1 − ‖y − Ĥx̂‖/‖y‖).

Input

Measurements y, ideal forward model H, spec ranges

Output

Reconstructed signal x̂

About the Imaging Modality

Multi-focus panoramic fusion combines images captured at different focal planes and/or different spatial positions to produce an all-in-focus image with extended depth of field and wide field of view. Focus stacking selects the sharpest regions from each focal plane using local contrast measures, then blends them via Laplacian pyramid fusion or wavelet-based methods. Panoramic stitching aligns overlapping images using feature matching (SIFT/SURF) and blends seams. Primary challenges include parallax at scene edges and focus measure ambiguity in low-texture regions.

Principle

Panoramic multi-focus fusion captures multiple images of the same wide scene at different focal distances and combines them to produce a single all-in-focus panorama with extended depth of field. Image stitching aligns overlapping frames using feature matching and homography estimation, while focus fusion selects the sharpest pixels from each focal plane.

How to Build the System

Mount a camera on a motorized panoramic head (nodal point rotation). For each pan/tilt position, capture a focus stack (3-10 images at different focus distances). Use a medium-aperture setting (f/5.6-f/8) for each frame. Stitch overlapping views (30 % horizontal overlap) and fuse focus stacks per view tile. Calibrate the panoramic head to rotate around the lens entrance pupil to minimize parallax.

Common Reconstruction Algorithms

  • Laplacian pyramid focus fusion (weighted blending by local contrast)
  • SIFT/SURF feature matching + RANSAC homography estimation
  • Multi-band blending (Burt-Adelson) for seamless stitching
  • Exposure fusion (Mertens et al.) for HDR panoramas
  • Deep-learning focus stacking (DFDF, DeepFocus)

Common Mistakes

  • Parallax errors from rotation not centered on the lens entrance pupil
  • Ghosting from moving objects between sequential captures
  • Color inconsistency between overlapping tiles due to auto-exposure variation
  • Incomplete focus coverage leaving blurry regions in the final panorama
  • Stitching artifacts at seam lines visible in the final output

How to Avoid Mistakes

  • Use a calibrated panoramic head; verify no-parallax point for the specific lens
  • Mask out or blend moving objects; capture quickly or use simultaneous multi-camera rigs
  • Lock exposure, white balance, and focus (manual mode) across all tiles
  • Plan focus distances to cover the entire depth range of the scene
  • Use multi-band blending and choose seam lines in textureless regions

Forward-Model Mismatch Cases

  • The widefield fallback applies Gaussian blur to a single image, but panoramic imaging involves geometric projection (cylindrical, spherical, or equirectangular) of the scene onto a wide field of view — the projection geometry is absent
  • Panorama multi-focus fusion requires modeling focus variation across the wide FOV and stitching multiple exposures — the widefield single-frame model cannot capture the spatially varying focus or overlap regions

How to Correct the Mismatch

  • Use the panorama operator that models the geometric projection (cylindrical or spherical warping) and focus-dependent blur across the wide field of view
  • Reconstruct using image stitching with homography estimation, exposure fusion, and spatially varying deblurring that account for the correct projection geometry

Experimental Setup — Signal Chain

Experimental setup diagram for Panorama Multi-Focus Fusion

Experimental Setup

Image Size: 4096x2048 (equirectangular)
Focus Planes: 6
Overlap Percent: 30
Fusion: Laplacian pyramid / wavelet fusion
Application: all-in-focus / extended depth of field

Key References

  • Burt & Adelson, 'The Laplacian Pyramid as a Compact Image Code', IEEE Trans. Commun. 31, 532-540 (1983)

Canonical Datasets

  • Lytro multi-focus test set

Spec DAG — Forward Model Pipeline

C(PSF_focus) → Σ_f → D(g, η₁)

C Depth-Dependent PSF (PSF_focus)
Σ Focus Stack Sum (f)
D Camera (g, η₁)

Mismatch Parameters

Symbol Parameter Description Nominal Perturbed
Δf focus_step Focus step error (μm) 0 2.0
Δr registration Inter-frame registration error (pixels) 0 0.5
ΔE exposure_variation Exposure variation (%) 0 3.0

Credits System

40%
Platform Profit Pool
Revenue allocated to benchmark rewards
30%
Winner Share
Top algorithm receives from pool
$100
Min Withdrawal
Minimum payout threshold
Spec Primitives Reference (11 primitives)
P Propagation

Free-space or medium propagation kernel (Fresnel, Rayleigh-Sommerfeld).

M Mask / Modulation

Spatial or spatio-temporal amplitude modulation (coded aperture, SLM pattern).

Π Projection

Geometric projection operator (Radon transform, fan-beam, cone-beam).

F Fourier Sampling

Sampling in the Fourier / k-space domain (MRI, ptychography).

C Convolution

Shift-invariant convolution with a point-spread function (PSF).

Σ Summation / Integration

Summation along a physical dimension (spectral, temporal, angular).

D Detector

Sensor readout with gain g and noise model η (Gaussian, Poisson, mixed).

S Structured Illumination

Patterned illumination (block, Hadamard, random) applied to the scene.

W Wavelength Dispersion

Spectral dispersion element (prism, grating) with shift α and aperture a.

R Rotation / Motion

Sample or gantry rotation (CT, electron tomography).

Λ Wavelength Selection

Spectral filter or monochromator selecting a wavelength band.