Skip to content

GATE2_5_RESULTS.md

Latest commit

 

History

History
95 lines (75 loc) · 2.42 KB

GATE2_5_RESULTS.md

File metadata and controls

95 lines (75 loc) · 2.42 KB
Raw

Gate 2.5 Results

Decision

  • Independent analytic fixture: PASSED
  • Device: cpu
  • Dtype: torch.float64
  • Reuses procedural renderer expected values: False
  • SC-GS modified before this gate: False

Checks

Check Result
camera_conventions PASS
finite_difference PASS
gradient_destinations PASS
invalid_inputs_rejected PASS
original_processed_transform PASS
perspective_projection_distinction PASS
projection_xy_origin_pixel_center PASS
se3_blend_then_project PASS
visibility_robust_losses_empty PASS

The fixture uses handwritten cameras, known camera-space points, explicit SE(3) matrices, fixed weights, known pixels, masks, and perturbations.

Numerical results

  • Camera transform max error: 0
  • Camera inverse round-trip max error: 0
  • Projection max error: 0 pixels
  • Half-pixel resize max error: 0 pixels
  • Canonical-anchor blend max error: 0
  • Deform-then-project max error: 0 pixels
  • Huber test loss: 4 over 2 valid observations
  • Empty-valid loss/count: 0 / 0

Finite-difference versus autograd:

  • autograd: -83.1758034026
  • numeric: -83.1758034323
  • absolute error: 2.9665258694e-08
  • relative error: 3.56657314625e-10
  • tolerances: abs 2e-06, relative 2e-05

Gradient destinations

  • Candidate/canonical anchor gradient norm: 185.592950584
  • Explicit rotation gradient norm: 255.82222789
  • Explicit translation gradient norm: 114.407178343
  • Association-score gradient present: False
  • Observation gradient present: False
  • Visibility gradient present: False
  • Discrete indices require gradients: False

Observed tracks, visibility, and discrete association are detached; gradients reach the selected canonical candidate and explicit deformation tensors.

Perspective distinction

project(sum_j w_j T_j X) = [25.0, 0.0]
sum_j w_j project(T_j X) = [50.0, 0.0]

Difference: 25 pixels.

This verifies that direct weighted node projection is not equivalent to the primary canonical-anchor-deform-then-project method.

Invalid inputs

NaN observations, Inf predictions, non-normalized weights, and out-of-range fixed associations all raised explicit ValueError exceptions. An empty visibility mask returns differentiable zero with valid count zero.

Gate

Gate 2.5 is PASSED. The pinned SC-GS reader may be audited next only if this result remains passed.