google-deepmind/orbit_sfm

google-deepmind/orbit_sfm

Language: Jupyter Notebook

License: Apache-2.0

Stars: 0

Forks: 0

Open issues: 0

Created: 2026-08-25T19:52:22Z

Pushed: 2026-08-25T21:03:05Z

Default branch: main

Fork: no

Archived: no

README:

ORBIT: A 360 Based Benchmark for evaluating SFM pipelines

Official code for "ORBIT".

Project Page | Paper

Installation

Follow the commands below to install the conda environment:

conda create -n orbit python=3.10 numpy scipy pandas matplotlib pillow \
scikit-image opencv jupyter
conda activate orbit
pip install evo

Provided Files and Folders

The project includes the following notebooks and data directories:

• `create_orbit.ipynb`: Notebook for generating benchmark clips by

projecting pre-processed panoramic images into perspective camera frames.

• `orbit_eval.ipynb`: Notebook for evaluating SfM methods against the

ground truth camera trajectories.

• `clip_lists.txt`: A text file listing all benchmark clips with their

source video ID, start time, and number of frames.

• `trajectories/`: Text files containing the camera intrinsics and extrinsic

rotations for each clip. Used by create_orbit.ipynb to render the evaluation clips.

• `cameras/`: Text files containing the ground truth camera paths (in TUM

format). Used by orbit_eval.ipynb to evaluate reconstructed trajectories.

Generating Benchmark Clips with create_orbit.ipynb

The input video should be pre-processed into a set of panoramic (equirectangular) images. The create_orbit.ipynb notebook then projects these panoramic images into the benchmark clips using the provided camera trajectories.

Before running, set the following variables in the notebook:

• images_dir: Path to the directory containing the source panoramic images,

organized by clip name.

• output_dir: Path to the directory where the projected clip images will be

saved.

The notebook reads clip_lists.txt to determine which clips to process and loads the corresponding trajectory files from the trajectories/ directory. For each clip, it reprojects the panoramic frames into perspective views and saves the resulting images to output_dir/{clip_name}/.

Evaluating SfM Methods with orbit_eval.ipynb

After running an SfM pipeline on the generated clips, use the orbit_eval.ipynb notebook to evaluate the reconstructed camera trajectories against the ground truth.

Before running, set the following configuration variables in the notebook:

• CAMERAS_DIR: Path to the cameras/ directory containing the ground truth

trajectories (default: './cameras').

• ESTIMATES_DIR: Path to the directory containing estimated trajectories,

organized as {method}/{clip_name}.txt (default: './estimates').

• METHODS: A list of method names to evaluate, corresponding to

subdirectories under ESTIMATES_DIR (e.g., ['colmap']).

The notebook computes ATE (Absolute Trajectory Error) and RPE (Relative Pose Error) metrics, plots per-clip trajectory comparisons, and generates summary tables and success-rate curves.

Example: Running with COLMAP

Installation

Please follow COLMAP's installation page to install COLMAP's CLI. The ORBIT codebase is tested on COLMAP 3.9.

Running

After saving the clip images using create_orbit.ipynb, run COLMAP as follows:

colmap feature_extractor --database_path [path to .db file] \
--image_path [path to images folder] \
--ImageReader.camera_model SIMPLE_RADIAL \
--SiftExtraction.use_gpu True \
--random_seed 0
colmap exhaustive_matcher --database_path [path to .db file] \
--SiftMatching.use_gpu True \
--random_seed 0
colmap mapper --database_path [path to .db file] \
--image_path [path to images folder] \
--output_path [results path]

Then place the estimated trajectories under ESTIMATES_DIR/colmap/ and run orbit_eval.ipynb to evaluate.