Skip to content

Performance improvement on generalized absolute pose.#4037

Merged
ahojnnes merged 2 commits into
mainfrom
shaohui/perf_gp3p
Jan 21, 2026
Merged

Performance improvement on generalized absolute pose.#4037
ahojnnes merged 2 commits into
mainfrom
shaohui/perf_gp3p

Conversation

@B1ueber2y

@B1ueber2y B1ueber2y commented Jan 20, 2026 •
edited
Loading

Copy link
Copy Markdown
Contributor

main (on the test case from #4023):

Running EstimateGeneralizedAbsolutePose with max_trials=1000
[RANSAC] 1000 trials, 2277 models, 58 inliers
  Estimation: 1.173 ms (11.5875%)
  Residuals:  8.005 ms (79.0773%)
  Scoring:    0.945 ms (9.33518%)
  Total:      10.123 ms
EstimateGeneralizedAbsolutePose RANSAC time: 0.010301 s, num_trials: 1000

After precomputing rig_from_world matrix:

Running EstimateGeneralizedAbsolutePose with max_trials=1000
[RANSAC] 1000 trials, 2277 models, 58 inliers
  Estimation: 1.14 ms (13.3834%)
  Residuals:  6.451 ms (75.7337%)
  Scoring:    0.927 ms (10.8828%)
  Total:      8.518 ms
EstimateGeneralizedAbsolutePose RANSAC time: 0.008678 s, num_trials: 1000

After precomputing cam_from_rig matrix:

Running EstimateGeneralizedAbsolutePose with max_trials=1000
[RANSAC] 1000 trials, 2279 models, 58 inliers
  Estimation: 1.196 ms (15.7431%)
  Residuals:  5.373 ms (70.7253%)
  Scoring:    1.028 ms (13.5317%)
  Total:      7.597 ms
EstimateGeneralizedAbsolutePose RANSAC time: 0.007778 s, num_trials: 1000

Will follow up another update on generalized relative pose.

Edit: In the generalized relative pose case poselib::gen_relpose_6pt is the bottleneck, so there is not much to optimize in the colmap side.

Edit: after switching to eigen expressions:

Running EstimateGeneralizedAbsolutePose with max_trials=1000
[RANSAC] 1000 trials, 2279 models, 58 inliers
  Estimation: 1.198 ms (18.0503%)
  Residuals:  4.455 ms (67.1237%)
  Scoring:    0.984 ms (14.826%)
  Total:      6.637 ms
EstimateGeneralizedAbsolutePose RANSAC time: 0.006816 s, num_trials: 1000

ahojnnes
ahojnnes reviewed Jan 20, 2026
View reviewed changes
Comment thread src/colmap/estimators/generalized_absolute_pose.cc Outdated
Comment thread src/colmap/estimators/generalized_absolute_pose.cc Outdated
@ahojnnes

ahojnnes commented Jan 20, 2026

Copy link
Copy Markdown
Contributor

An alternative optimization could be to store the "camera_id" and then precompute the cam_from_world matrix once and reuse it for direct transformation from world to camera instead of doing two transformations. Probably depends a bit on the problem and how many cameras it contains on whether this is faster.

@ahojnnes ahojnnes merged commit 754b099 into main Jan 21, 2026
14 checks passed
@ahojnnes ahojnnes deleted the shaohui/perf_gp3p branch January 21, 2026 12:15
@nushakrishnan nushakrishnan mentioned this pull request May 5, 2026
Merged
ahojnnes added a commit that referenced this pull request May 5, 2026
@nushakrishnan @gemini-code-assist @ahojnnes
Fix: Missing rig pose manifold in RefineGeneralizedAbsolutePose (#4375)
4bc00e9 
Hi

I was facing an error with
`pycolmap.estimate_and_refine_generalized_absolute_pose` when installing
pycolmap via pip.

Error:
```
*** Aborted at 1777983489 (unix time) try "date -d @1777983489" if you are using GNU date ***
PC: @     0x73ded64220cb (/local/home/akrishnan/lamar_env_fix/lib/python3.10/site-packages/pycolmap/_core.cpython-310-x86_64-linux-gnu.so+0x10220ca)
Segmentation fault
```

The breaking PR was #4041. I was able to validate this by installing
pycolmap from source on the commit before (#4037). The script below
should work on PR #4037.

Test data:
[Rig_pose_data](https://drive.google.com/file/d/125Y-Pap9FhQUVCayLVfU6iTWXQSHJaS7/view?usp=sharing)

Testing script:

```python
import pickle, time
import numpy as np
import pycolmap
from tqdm import tqdm

PKL = '/local/home/akrishnan/Downloads/rig_pose_debug.pkl'

with open(PKL, 'rb') as f:
    all_data = pickle.load(f)

print(f'pycolmap {pycolmap.__version__} from {pycolmap.__file__}')
print(f'{len(all_data)} entries')

n_ok = 0
n_none = 0
t0 = time.time()
for i, d in enumerate(tqdm(all_data, desc='rig PnP')):
    p2d_per_cam = d['p2d']
    p3d_per_cam = d['p3d']
    camera_dicts = d['camera_dicts']
    qvecs = d['qvecs'] 
    tvecs = d['tvecs']
    thresh = float(d['thresh'])

    p2d_flat = np.concatenate(p2d_per_cam, axis=0).astype(np.float64)
    p3d_flat = np.concatenate(p3d_per_cam, axis=0).astype(np.float64)
    camera_idxs = []
    for ci, p in enumerate(p2d_per_cam):
        camera_idxs.extend([ci] * len(p))

    cams_from_rig = []
    for q_wxyz, t in zip(qvecs, tvecs):
        q_xyzw = np.array([q_wxyz[1], q_wxyz[2], q_wxyz[3], q_wxyz[0]], dtype=np.float64)
        cams_from_rig.append(pycolmap.Rigid3d(
            pycolmap.Rotation3d(q_xyzw),
            np.asarray(t, dtype=np.float64),
        ))

    # Build cameras
    cameras = [
        pycolmap.Camera(
            model=cd['model'],
            width=int(cd['width']),
            height=int(cd['height']),
            params=list(cd['params']),
        )
        for cd in camera_dicts
    ]

    opts = pycolmap.RANSACOptions()
    opts.max_error = thresh

    ret = pycolmap.estimate_and_refine_generalized_absolute_pose(
        p2d_flat, p3d_flat, camera_idxs, cams_from_rig, cameras,
        estimation_options=opts, return_covariance=True,
    )
    if ret is None:
        n_none += 1
    else:
        n_ok += 1

dt = time.time() - t0
print(f'\ndone in {dt:.1f}s — succeeded: {n_ok}/{len(all_data)}, '
      f'returned None: {n_none}/{len(all_data)}')
```

---------

Co-authored-by: gemini-code-assist[bot] <176961590+gemini-code-assist[bot]@users.noreply.github.com>
Co-authored-by: Johannes Schönberger <jsch@demuc.de>
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Reviewers

@ahojnnes ahojnnes ahojnnes approved these changes
@sarlinpe sarlinpe Awaiting requested review from sarlinpe

Assignees

No one assigned

Labels

None yet

Projects

None yet

Milestone

No milestone

Development

Successfully merging this pull request may close these issues.

2 participants

@B1ueber2y @ahojnnes