P
Physics World Model
Coded Aperture Compressive Temporal Imaging (CACTI) (L1-004)

CACTI Mismatch-Only Spec (standard SCI video benchmark family) L2-004

mismatch_onlyd_spec = 0.45★ Founder-vetted

Six-tuple S = (Ω, E, B, I, O, ε)

Ω — parameter space

{
  "H": [
    32,
    1024
  ],
  "W": [
    32,
    1024
  ],
  "B": [
    4,
    32
  ],
  "noise_level": [
    0.001,
    0.1
  ],
  "dx": [
    0,
    1
  ],
  "dy": [
    0,
    1
  ],
  "theta": [
    0,
    0.15
  ],
  "dt": [
    0,
    0.1
  ],
  "eta": [
    0.9,
    1
  ],
  "gamma": [
    0.95,
    1.05
  ],
  "offset": [
    0,
    0.01
  ],
  "sigma_n": [
    0.5,
    2
  ]
}

E — forward operator

Primitive chain:
S.temporal.coded -> int.temporal

Inverse: recover x in R^{H x W x B} from a single 2D snapshot y

B — boundary constraints

  • nonnegativity: true
  • temporal_continuity: true
  • mask_binary: C_b(i,j) in {0,1}

I — initialization

zero_init

O — observable outputs

per_frame_PSNRSSIMtemporal_consistencyresidual_norm

ε — acceptance threshold

26.0 + 1.5 * log2(B / 8) + 1.0 * log10(photon_count / 100)

Protocol fields

Input format

{
  "measurement": "float32(H, W)",
  "mask": "bool(H, W, B)"
}

Output format

{
  "video_cube": "float32(H, W, B)"
}

Authored baselines

SolverMethod sigExpected metric
GAP-TVL+Oexpected_psnr=26.5
DeSCIL+Oexpected_psnr=28.3
PnP-FFDNetL+Nexpected_psnr=30.1
EfficientSCIT+Nexpected_psnr=34.2
BIRNATR+Nexpected_psnr=33

Child benchmarks (1)