Repository navigation
Add Depth-Anything-3 reference implementation - #44
Open
zitongzhan wants to merge 8 commits into
Open
zitongzhan wants to merge 8 commits into
zitongzhan wants to merge 8 commits into
Conversation
![github-code-quality[bot]](https://avatars.githubusercontent.com/u/9919?s=60&v=4){kind=small}
Comment on lines
+147
to
+149
| 'All methods receive the same globally attended, undistorted DSLR prediction for each scene. ' | ||
| '`BAE_USE_PYPOSE_AMBIENT_GRAD=1` is enabled. No scene is selected using its AUC. ' | ||
| 'Raw and guarded outcomes are separate. All table AUCs are percentages; JSON stores fractions.', '', |
Comment on lines
+160
to
+162
| lines += ['', 'Native D minimizes a joint photometric objective; Open3D D performs dense pairwise odometry followed by PGO. ' | ||
| 'Native E rebuilds a voxel-averaged structure; reference E uses TSDF fusion and rigid color-map optimization. ' | ||
| 'These are different formulations, so numerical equality is not guaranteed by correct gradients.', '', |
Comment on lines
+177
to
+180
| '- The reference arms here call the actual DA3 Form D/E functions on the same cached arrays as bae. ' | ||
| 'They use the showcase protocol: shared self-calibration, graph baseline filter 0.1 / rotation 1°, ' | ||
| 'D texture gate and internal guard disabled, then a common 512-pair wide guard. ' | ||
| 'D+E composes raw D then E and guards the whole proposal. This differs from the historical CLI defaults.', |
Comment on lines
+185
to
+186
| '- Camera ground truth is used only for evaluation. No optimizer receives COLMAP poses or calibrated intrinsics. ' | ||
| 'Fixed DSLR calibration is used only to undistort the input photographs.', |
Comment on lines
+187
to
+188
| '- Per-scene input filenames, memory search attempts, checkpoint revision, configuration and metrics are in JSON. ' | ||
| 'Full pair errors are saved beside each prediction as `evaluation_errors.npz`.', '', |
| independent Ceres-style reference here, so we pin to "recovers a known | ||
| synthetic ground truth" instead of a magic reference cost). | ||
| """ | ||
| import os |
| sys.path.insert(0, p) | ||
|
|
||
| import pypose as pp | ||
| from pypose.autograd.function import psjac |
|
|
||
| from bae.autograd.graph import jacobian | ||
|
|
||
| from geometry import as_44, affine_inv, unproject_pixels, project_to_pixels |
This branch has not been deployed
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
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 pull request introduces a comprehensive RGB-D pose refinement example and integrates a benchmarking and validation protocol for comparing pre- and post- refinement result.
The most important changes are:
Documentation and Protocols:
README.md,SHOWCASE.md, andVALIDATION.mdwith detailed instructions for running, reproducing, and validating DSLR pose refinement experiments, including dependencies, input requirements, evaluation protocols, and gradient correctness checks. [1] [2] [3]benchmark/README.mdsummarizing the purpose and usage of benchmark scripts.Benchmarking and Evaluation Code:
benchmark/showcase_benchmark.py, a script that runs and compares native BAE odometry/structure refinement with DA3 reference implementations, manages caching, reproducibility, and reports detailed metrics for each variant.