CASSI

Coded Aperture Snapshot Spectral Imaging

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
🥇 MiJUN-5stg 0.927 40.9 0.991 ✓ Certified Meng et al. AAAI 2025
🥈 RDLUF-MixS2-9stg 0.904 39.6 0.988 ✓ Certified Dong et al. CVPR 2023
🥉 DAUHST-9stg 0.883 38.4 0.985 ✓ Certified Cai et al. NeurIPS 2022
4 PADUT-3stg 0.854 36.95 0.975 ✓ Certified Li et al. ICCV 2023
5 CST-L-Plus 0.836 36.1 0.967 ✓ Certified Cai et al. ECCV 2022
6 MST++ 0.833 36.0 0.966 ✓ Certified Cai et al. CVPRW 2022
7 MST-L 0.809 34.81 0.958 ✓ Certified Cai et al. CVPR 2022
8 HDNet 0.804 34.66 0.952 ✓ Certified Hu et al. CVPR 2022
9 SSR-L 0.797 34.0 0.960 ✓ Certified Zhang et al. CVPR 2024
10 DGSMP 0.752 32.6 0.917 ✓ Certified Huang et al. CVPR 2021
11 TSA-Net 0.722 31.5 0.894 ✓ Certified Meng et al. ECCV 2020
12 λ-Net 0.696 30.1 0.887 ✓ Certified Miao et al. ICCV 2019
13 BIRNAT 0.694 30.0 0.887 ✓ Certified Cheng et al. ECCV 2022
14 ADMM-Net 0.674 29.1 0.877 ✓ Certified Ma et al. ICCV 2019
15 GAP-Net 0.669 29.1 0.867 ✓ Certified Meng et al. 2020
16 GAP-TV 0.577 24.34 0.820 ✓ Certified Yuan et al. 2016
17 PnP-HSICNN 0.573 25.12 0.810 ✓ Certified Maffei et al. 2020
18 TwIST 0.538 23.1 0.800 ✓ Certified Bioucas-Dias & Figueiredo 2007

Dataset: KAIST simu, 256×256×28

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
🥇 SSR-L + gradient 0.626
0.877
38.03 dB / 0.994
0.545
19.53 dB / 0.626
0.456
17.06 dB / 0.473
✓ Certified PWM benchmark (CVPR 2024)
🥈 GAP-TV + gradient 0.593
0.687
24.21 dB / 0.865
0.576
19.69 dB / 0.699
0.516
18.37 dB / 0.583
✓ Certified PWM benchmark
🥉 MST-L + gradient 0.550
0.794
31.29 dB / 0.977
0.472
17.18 dB / 0.550
0.385
15.45 dB / 0.384
✓ Certified PWM benchmark
4 PnP-HSICNN + gradient 0.504
0.549
20.06 dB / 0.675
0.508
16.88 dB / 0.621
0.455
15.59 dB / 0.518
✓ Certified PWM benchmark
5 HDNet + gradient 0.439
0.707
25.45 dB / 0.921
0.329
13.59 dB / 0.427
0.280
10.96 dB / 0.373
✓ Certified PWM benchmark

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 10 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 SSR-L + gradient 0.877 38.03 0.994
2 MST-L + gradient 0.794 31.29 0.977
3 HDNet + gradient 0.707 25.45 0.921
4 GAP-TV + gradient 0.687 24.21 0.865
5 PnP-HSICNN + gradient 0.549 20.06 0.675
Spec Ranges (5 parameters)
Parameter Min Max Unit
mask_dx 0.3 0.7 px
mask_dy 0.1 0.5 px
mask_rotation 0.0 0.2 deg
dispersion_slope 1.895 2.145 px/band
dispersion_axis 0.0 0.3 deg
View Details →
Dev 10 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 GAP-TV + gradient 0.576 19.69 0.699
2 SSR-L + gradient 0.545 19.53 0.626
3 PnP-HSICNN + gradient 0.508 16.88 0.621
4 MST-L + gradient 0.472 17.18 0.55
5 HDNet + gradient 0.329 13.59 0.427
Spec Ranges (5 parameters)
Parameter Min Max Unit
mask_dx 0.4 0.8 px
mask_dy 0.2 0.6 px
mask_rotation 0.05 0.25 deg
dispersion_slope 1.825 2.075 px/band
dispersion_axis 0.07 0.37 deg
Submit Reconstruction View Details →
Hidden 10 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 GAP-TV + gradient 0.516 18.37 0.583
2 SSR-L + gradient 0.456 17.06 0.473
3 PnP-HSICNN + gradient 0.455 15.59 0.518
4 MST-L + gradient 0.385 15.45 0.384
5 HDNet + gradient 0.280 10.96 0.373
Spec Ranges (5 parameters)
Parameter Min Max Unit
mask_dx 0.2 0.6 px
mask_dy 0.0 0.4 px
mask_rotation -0.05 0.15 deg
dispersion_slope 1.955 2.205 px/band
dispersion_axis -0.05 0.25 deg
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

CASSI captures a 3D hyperspectral data cube (2 spatial + 1 spectral dimension) in a single 2D camera exposure. The scene is modulated by a binary coded aperture mask, spectrally dispersed by a prism, and integrated onto a 2D detector. The forward model is y = H*x + n where H encodes both coded-aperture modulation and spectral-dispersion shift. Compression ratios equal the number of spectral bands (e.g. 28:1). Reconstruction exploits spectral correlation via GAP-TV, MST, or CST.

Principle

Coded Aperture Snapshot Spectral Imaging (CASSI) captures a full 3-D spectral datacube (x, y, λ) in a single 2-D snapshot by encoding the scene with a binary coded aperture and spectrally dispersing it with a prism onto the detector. Different spectral channels are shifted and superimposed on the sensor, creating a compressed measurement. Computational algorithms recover the full datacube from this single measurement using sparsity priors.

How to Build the System

Build an optical relay with an objective lens, place a binary coded aperture (lithographic chrome-on-glass mask or DMD) at an intermediate image plane, then disperse with an Amici or double-Amici prism, and re-image onto a high-resolution detector (2048× 2048+ pixels). Precisely calibrate the spectral dispersion curve (nm/pixel). The coded aperture pattern should have ~50 % transmittance and good conditioning.

Common Reconstruction Algorithms

  • TwIST (Two-step Iterative Shrinkage/Thresholding)
  • GAP-TV (Generalized Alternating Projection with Total Variation)
  • ADMM with sparsity in DCT or wavelet domain
  • Deep unfolding networks (DGSMP, TSA-Net, BIRNAT)
  • Plug-and-Play ADMM with learned denoisers

Common Mistakes

  • Poor spectral calibration causing wavelength assignment errors across the datacube
  • Coded aperture not precisely at the image plane, blurring the code modulation
  • Insufficient detector resolution relative to the number of spectral bands
  • Ignoring optical aberrations in the dispersive relay that vary with wavelength
  • Using a random mask without checking its sensing matrix condition number

How to Avoid Mistakes

  • Calibrate spectral mapping with monochromatic sources at known wavelengths
  • Mount coded aperture on a precision z-stage and focus to maximize modulation contrast
  • Ensure detector pixel count > (spatial pixels × spectral bands) for adequate compression ratio
  • Design the relay optics for uniform imaging quality across the spectral range
  • Optimize or simulate the mask pattern for low coherence (good RIP) before fabrication

Forward-Model Mismatch Cases

  • The widefield fallback produces a 2D (64,64) grayscale image, but CASSI compresses a 3D spectral datacube (64,64,L wavelengths) into a single 2D coded snapshot via a binary mask and dispersive prism — the spectral dimension is entirely absent
  • Without the coded aperture mask and spectral dispersion, the measurement does not encode wavelength-dependent information — spectral unmixing or hyperspectral reconstruction from the fallback output is impossible

How to Correct the Mismatch

  • Use the CASSI operator that applies the binary coded aperture mask followed by spectral dispersion (prism/grating shift), producing a 2D coded measurement that encodes the full 3D spectral datacube
  • Reconstruct the (x,y,lambda) datacube using compressive sensing (TwIST, GAP-TV) or deep unfolding networks (TSA-Net, MST) that exploit the spatio-spectral structure encoded by the CASSI forward model

Experimental Setup — Signal Chain

Experimental setup diagram for Coded Aperture Snapshot Spectral Imaging (CASSI)

Experimental Setup

Instrument: Custom SD-CASSI / KAIST CASSI prototype
Coded Aperture: binary random mask on photolithography substrate
Disperser: Amici prism (SD-CASSI)
Spectral Bands: 28
Wavelength Range Nm: 450-650
Spatial Resolution: 256x256
Compression Ratio: 28
Detector: FLIR Grasshopper3 monochrome CMOS (2048x2048)
Relay Lens: 4f relay system with 1:1 magnification

Key References

  • Wagadarikar et al., 'Single disperser design for coded aperture snapshot spectral imaging', Applied Optics 47, B44-B51 (2008)
  • Cai et al., 'Mask-guided Spectral-wise Transformer (MST++)', CVPRW 2022

Canonical Datasets

  • CAVE (Columbia, 32 scenes, 512x512x31)
  • KAIST (30 scenes, 2704x3376x28)
  • ARAD_1K (1000 hyperspectral images)

Spec DAG — Forward Model Pipeline

M(mask) → W(α, a) → Σ_λ → D(g, η₄)

M Coded Aperture (mask)
W Prism Dispersion (α, a)
Σ Spectral Sum (λ)
D Detector (g, η₄)

Mismatch Parameters

Symbol Parameter Description Nominal Perturbed
Δx mask_dx Mask lateral shift (pixels) 0 0.5
Δy mask_dy Mask vertical shift (pixels) 0 0.3
θ mask_theta Mask rotation (rad) 0 0.1
a₁ disp_a1 Dispersion coefficient 2.0 2.02
α disp_alpha Dispersion angle (rad) 0 0.15
σ_r sigma_read Detector read noise std (electrons) 5.0 8.0
I_d dark_current Dark current (electrons/pixel/s) 0.1 0.5
g gain Detector gain multiplier 1.0 1.03

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.