Debian Codegraph @

analyses/splines/README

raw
1 The test and validation tools used in the development of the b-spline
2 interpolated models live here.
3
4 First I wanted to come up with a surface parametrization scheme. I wanted to
5 project the full almost-360deg view into some sort of 2D domain that is
6 cartesian-ish. Then I would grid this cartesian 2D domain with a regular grid of
7 control points, and I'd spline that. But does such a mapping exist? What is it?
8
9 The
10
11 show-fisheye-grid.py
12
13 script exists for this purpose. It takes in an existing model, grids the imager,
14 unprojects the points, and projects the vectors back to a number of common
15 fisheye projections. I applied this to some wide models I had (their inaccurate
16 OPENCV8 fits), and observed that the stereographic model works well here: a
17 cartensian grid in the imager produces a cartesian-ish grid in the stereographic
18 projection. The other projections look like they'd work also. The pinhole
19 projection most notably does NOT work; its projections run off to infinity at
20 the edges.
21
22 Then I needed to find a surface representation, and the expressions. I'm using
23 b-splines: they have nice continuity properties, and they have nice local
24 support: so my jacobian will remain sparse. The
25
26 bsplines.py
27
28 script derives and validates the spline equations. fpbisp.py is a part of that.
29
30 Then I implemented the spline equations in mrcal.c, and wanted to make sure the
31 behavior was correct. This is done by the
32
33 verify-interpolated-surface.py
34
35 script. It produces a random splined model, projects it with mrcal, and
36 visualizes the sparse control point surface with the dense sampled surface. The
37 sampled surface should appear smooth, and should be guided by the control
38 points. This tool is the first test that the mrcal projection is implemented
39 correctly