Issues

[Bug]: So3.log, the So3 Jacobians and Se3.exp/log lose all their digits for small rotations (float32 returns 0 below 1e-4 rad, Jacobians are nan at the identity)

#4897

· Nicholas022400701 opened

on Sep 25, 2026

`find_homography_dlt`: a correspondence with weight 0 still changes H through the point normalisation (0.5–1 px with 1-px noise)

#4890

· ducha-aiki opened

on Sep 24, 2026

tests: `generate_two_view_random_scene` is unseeded, so `test_from_to_essential[cpu-bfloat16]` and its other callers are flaky outside the seeded CI profiles

#4886

· ducha-aiki opened

on Sep 24, 2026

`find_essential` in float32 misses the true essential matrix on an exact 5-point sample (N = 5 only; N >= 6 is fine)

#4884

· ducha-aiki opened

on Sep 24, 2026

`find_essential` returns ten identity matrices when a 5-point sample has no real solution, while complex slots are otherwise NaN

#4883

· ducha-aiki opened

on Sep 24, 2026

Epipolar and homography error metrics add `eps` to the value: Sampson 90% low at ‖F‖ = 1e-3, exact matches score 1e-4

#4881

· ducha-aiki opened

on Sep 24, 2026

`decompose_essential_matrix_no_svd` returns non-rotations for any batch of two or more, even `[E, E]`

#4880

· ducha-aiki opened

on Sep 24, 2026

`motion_from_essential_choose_solution` silently returns candidate 0 when no point passes the depth test

#4879

· ducha-aiki opened

on Sep 24, 2026

`decompose_essential_matrix` / `motion_from_essential` batch the rotations of a `(3, 3)` input but not `t`

#4878

· ducha-aiki opened

on Sep 24, 2026

`fundamental_from_projections` overflows to `inf` in float16 for pixel-unit projection matrices (5 of 9 entries)

#4877

· ducha-aiki opened

on Sep 24, 2026

`find_essential` documents per-correspondence `weights` but ignores them: all-zero weights change nothing

#4876

· ducha-aiki opened

on Sep 24, 2026

`find_fundamental`: a correspondence with weight 0 still changes F through the Hartley normalisation

#4875

· ducha-aiki opened

on Sep 24, 2026

Navigation Menu

Uh oh!

All issues

[Bug]: So3.log, the So3 Jacobians and Se3.exp/log lose all their digits for small rotations (float32 returns 0 below 1e-4 rad, Jacobians are nan at the identity)

`find_homography_dlt`: a correspondence with weight 0 still changes H through the point normalisation (0.5–1 px with 1-px noise)

tests: `generate_two_view_random_scene` is unseeded, so `test_from_to_essential[cpu-bfloat16]` and its other callers are flaky outside the seeded CI profiles

`find_essential` in float32 misses the true essential matrix on an exact 5-point sample (N = 5 only; N >= 6 is fine)

`find_essential` returns ten identity matrices when a 5-point sample has no real solution, while complex slots are otherwise NaN

Epipolar and homography error metrics add `eps` to the value: Sampson 90% low at ‖F‖ = 1e-3, exact matches score 1e-4

`decompose_essential_matrix_no_svd` returns non-rotations for any batch of two or more, even `[E, E]`

`motion_from_essential_choose_solution` silently returns candidate 0 when no point passes the depth test

`decompose_essential_matrix` / `motion_from_essential` batch the rotations of a `(3, 3)` input but not `t`

`fundamental_from_projections` overflows to `inf` in float16 for pixel-unit projection matrices (5 of 9 entries)

`find_essential` documents per-correspondence `weights` but ignores them: all-zero weights change nothing

`find_fundamental`: a correspondence with weight 0 still changes F through the Hartley normalisation

Uh oh!

Search results