Introduces unproject() lookup tables#29
Open
csp256 wants to merge 18 commits into
Open
Conversation
Owner
|
Thanks for the PR, it's a great addition. I'm looking over it and making some changes, I'll get back when I've gone through it properly. |
Replace the Python adaptive-quadtree analyzer with a per-cell C++ optimiser that maximises sin²(angular error) over each LUT cell using gradient ascent in normalised camera-frame coordinates with a ReLU penalty enforcing pixel-cell membership. Backed by ceres::Jet autodiff, OpenMP across cells, and works for both PinholeSplined and OpenCV models. The heatmap and accuracy report both use this single C++ entry. The dense sampling path (sample_lut_accuracy / UnprojectLUTSampleAccuracy) is removed — the per-cell maximiser supersedes it. Heatmap computation goes from ~6s to ~0.2s on typical LUTs. Angular error formula switched to atan2(‖cross‖, dot) for full precision near zero. Adds heatmap correctness tests covering optimiser/analyzer agreement, peak-pixel reproduction, cell-membership, delta consistency, brute-force matching on a realistic LUT, and exact zero error for a linear pinhole. Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
Restructure UnprojectLUTErrorHeatmap to store the three primary 2D arrays (peak_pixel_xy, exact_xy, approx_xy) and expose max_angular_error_deg, error_delta_xy, and error_direction_xy as @Property derivations. The C++ output column layout changes from (N, 5) [angular, peak, delta] to (N, 6) [peak, exact, approx]; Python derives the rest. Save format drops 3 redundant entries. Also drop max_iters / grad_tol from the public Python API and from both result dataclasses — they're now hardcoded as private constants (_OPTIMISER_MAX_ITERS, _OPTIMISER_GRAD_TOL). Users never need to tune them in practice. Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
Add `vmax` parameter for capping the colorbar in the chosen angular_unit. Default arrow length now scales with per-cell error magnitude (brightest cell maps to `arrow_scale * arrow_spacing`); pass `constant_arrow_length=True` for the previous fixed-length behaviour. Drop the `Path | str` overload from the heatmap argument — callers always pass a heatmap object now. Trims the unused Path import. Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
Reject grids smaller than 2x2 or with zero-span extents in the constructor instead of papering over them with scale=0 branches. Move each interpolation mode into its own private method so the dispatch in query() is a plain switch. Replace the string_view + char-storage trick with std::string for lensboy_version, and make the invalid-result path explicit via UnprojectLUTQueryResult::invalid(). Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
- Wgrid/Hgrid -> grid_width/grid_height (lut_max_cell_error.cpp) - nx/ny family -> normalized_x/normalized_y across 4 files - px/py family -> pixel_x/pixel_y across all .cpp files; py:: namespace preserved - Drop > 1 defensive guards already guaranteed by entry-point require >= 2 (bilinear interpolator, scale + pixel-span calculations) - Indent OpenMP pragmas with // clang-format off/on scaffolding Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
2 participants
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
This introduces the ability to precompute and save to disk a sparse (or dense) lookup table of
normalize_points()results. It also introduces a way to analyze the accuracy of this lookup table.There is a
cpp_runtime/folder, which is intended to be copy-pasted into the user's C++ project to read and use the.unproject_LUTfile. This code in this folder is not connected to any of the extantcpp_src/code.I added tests and they all pass.
Building a dense lookup table can take quite a while on a large image, so that part is parallelized. The analysis code can still be quite slow, but it's usable and was resistant to initial efforts to parallelize it. I stopped myself from lowering that to C++.
I added docs, as well as an example Marimo notebook, which you can run by simply going
./example/unproject_lut_mo.py. The_mosuffix on the stem enables Jupyter to recognize it as a marimo notebook with https://github.com/marimo-team/marimo-jupyter-extension This is a bit of what it looks like:Maybe you want to convert this to a Jupyter notebook? I don't even have Jupyter installed any more. It might be a good idea to provide examples as both Jupyter and Marimo notebooks. I can open another PR to convert the Jupyter notebooks to Marimo, if you'd like.
I moved the contents of CLAUDE.md to AGENTS.md and pointed CLAUDE.md at it, as I understand is the standard in projects where people use multiple coding agents.
I know its not radically simple, but hopefully it doesn't violate the KISS too much for your taste.
Some of the old logic about previous file format versions has stuck around, but I can clean that up on Monday. I just wanted to get this PR to you before the weekend.