CASSI Mismatch-Only Benchmark Suite (P-benchmark + 4 I-benchmark tiers)L3-003

ReconstructMismatch

↑ L1-003 ↑ L2-003

Read full 4-layer walkthrough →

Example Data

View demos →

A reference solver demo is available — see input / ground truth / reconstruction visualizations on the demos page.

Genesis spec

Parent L2:: L2-003
Spec type:: —
d_spec:: —

On-chain

Not registered on-chain yet.

P-benchmark — rho=50 distribution over Ω

The headline ranking score. Solvers are evaluated on 50 instances drawn parametrically across the full Ω, scored as a distribution. This is what the leaderboard's rank-1 chases.

Dataset: KAIST-30 + CAVE + ICVL union with parametric Omega sampling(200 dev instances)

Authored baseline	overall_Q	≈ Q_int
GAP-TV	0.46	46
ADMM-CASSI	0.58	58
PnP-HSICNN	0.7	70

To take rank 1 of the P-benchmark, beat the highest authored Q above. See L2 spec for the epsilon_fn used per draw.

I-benchmark tiers — 4 fixed-Ω operating points

Each tier is a single-instance evaluation at a specific Ω with its own ε floor. Solvers may pass individual tiers without saturating the full P-benchmark.

Tier	ρ	ε (dB)	Ω parameters	d_ibench	Top authored baseline
T1_nominal	1	28	H=256, W=256, N_bands=28 …	0.15	PnP-HSICNN = 31.8 (PSNR_dB)Q=0.89
T2_low	3	25.5	H=256, W=256, N_bands=28 …	0.32	PnP-HSICNN = 28.3 (PSNR_dB)Q=0.74
T3_moderate	5	23	H=512, W=512, N_bands=28 …	0.51	PnP-HSICNN = 25.4 (PSNR_dB)Q=0.64
T4_blind	10	21.5	H=1024, W=1024, N_bands=28 …	0.72	PnP-HSICNN = 22.5 (PSNR_dB)Q=0.52

Current standing

🏆 Current SOTA

No certs submitted yet.

Be the first to submit a solution and claim rank 1.

📊 Classical floor

GAP-TV

26.00 dB

deliberate floor; anyone better wins

🧠 Deep-learning floor

MST-L

35.30 dB

published deep-learning landmark; harder gate

📈 PWM-enabled delta

Delta unavailable (need both SOTA and reference PSNR).

Example data

Two independent samples from the InverseNet benchmark, each shown as measurement → ground truth → reference reconstruction. The reconstruction is the reference solver's output — the bar a competing solver must beat to win the bounty.

sample_01 · scene01

reference PSNR 35.3 dB

Measurement

shape [256,310]

Ground truth

shape [256,256,28]

Reference reconstruction

35.3 dB

shape [256,256,28]

sample_02 · scene02

reference PSNR 36.13 dB

Measurement

shape [256,310]

Ground truth

shape [256,256,28]

Reference reconstruction

36.13 dB

shape [256,256,28]

Get this benchmark

Everything you need to run the reference solver locally, browse the full benchmark specification, or submit a competing solution.

Download sample data

sample_01 · 13487860 B

sample_02 · 13711391 B

sample_03 · 13598392 B

sample_04 · 13438897 B

sample_05 · 13509118 B

sample_06 · 13575530 B

sample_07 · 13389948 B

sample_08 · 13626068 B

sample_09 · 13542125 B

sample_10 · 13553346 B

Browse on GitHub

Benchmark card (YAML spec)
L3 principle JSON (canonical on-chain spec)
Reference solver (Python)
Demo README (full pipeline walkthrough)
Bounty index (compete with a better solver)

Run the reference solver locally

git clone https://github.com/integritynoble/pwm-public.git
cd pwm-public
python3 scripts/regen_demos_mst_l.py  (uses pwm_core.recon.mst.mst_recon_cassi with mst_l weights)
cat /tmp/out/meta.json

Expected reference PSNR: 35.3 dB (seed ). Byte-stable across runs at the same git SHA.

Leaderboard (top 10)

No solutions submitted yet. The reference baseline above is the floor — submit a better solver to claim rank 1.