L3 Benchmark ⊙ Testnet L3-003-T1

CASSI

Coded Aperture Snapshot Spectral Imaging

🚀 Submit a solution — mine PWM View benchmark MD →
Lineage: L1 Principle · L1-003 → L2 Spec · L2-003 → L3 CASSI Benchmark

How the CASSI forward model works

A binary coded aperture modulates the scene, a prism disperses each wavelength band by a pixel-linear shift, and the detector integrates the modulated, dispersed cube onto a single 2-D snapshot. Recovering the full hyperspectral cube from one snapshot is the inverse problem.

3-D cube
f(x,y,λ)
→
mask C
(binary)
→
prism shear
shift a·λ
→
Σ over λ
(detector)
→
2-D snapshot
y(x,y)

Standard benchmark

nominal · no mismatch

One benchmark, evaluated on the full dataset — no separate public / hidden / dev splits. KAIST-30 center 256x256 crop, 28 bands at 10 nm spacing across 450-650 nm; calibrated mask and dispersion (no mismatch).

28.0

Success threshold ε

256×256

Resolution

28

Spectral bands

3

Reference baselines

Reference baselines shipped with the benchmark

Solver Quality Q Status
baseline:PnP-HSICNN 0.890 ⊙ testnet
baseline:ADMM-CASSI 0.760 ⊙ testnet
baseline:GAP-TV 0.750 ⊙ testnet

Community leaderboard 3 submitted

# Solver Quality Q Status
1 real:GAP-Net@v1.0 1.000 ⊙ testnet
2 real:MST-L@v1.0 1.000 ⊙ testnet
3 real:ADMM-Net@v1.0 0.965 ⊙ testnet
⚠

Mismatch is its own benchmark + spec

The standard benchmark above assumes a perfectly-calibrated system (all mismatch parameters = 0). Real sensors have calibration error — mask shift, rotation, dispersion-slope drift. Those are tested under a separate mismatch spec and benchmark, where Ω includes the mismatch dimensions (mask_dx, mask_dy, mask_theta, disp_a1_error, disp_alpha_error).

Mismatch Spec · L2-003 → Mismatch benchmark tiers →