Docs
Pipelines
Besides the model families, fastforks ships a few pipelines that chain models and geometry on the GPU. They are not entries in the model table; this page is where to find them.
Structure from motion
An image folder becomes a COLMAP database, optionally a sparse reconstruction: image retrieval picks the pairs, a matcher finds correspondences, a GPU RANSAC verifies them and pycolmap maps the result. Matcher choice, speed and accuracy on Mip-NeRF 360 and ETH3D are in the guide.
fastforks sfm scene/images scene/out --reconstruct Calibrated stereo and rectification
Raw, possibly distorted pairs and a calibration file (OpenCV, KITTI, OAK-D, ZED, Aria and more) go in; a stereo model runs on the rectified pair and the result comes back as depth and a point cloud at the original left pixels. The rectification itself (camera models, agreement with OpenCV, speed) is a separate module.
fastforks stereo run left/ right/ --model fast-acvnet-generalization --rig calib.yaml --out depth/ Stereo geometry and rectification → · Calibrated stereo in one call →
Two-view and absolute pose (GPU RANSAC)
Matches from any matcher become a relative pose (5-point essential matrix), a
fundamental matrix (7-point) or an absolute pose from 2D-3D matches (P3P), with batched RANSAC following
OpenCV's protocol. Plain PyTorch or one CUDA kernel library (mode="kernels"). The same RANSAC
verifies the pairs in the structure-from-motion pipeline.
from fastforks.matching import estimate_pose
pose = estimate_pose(points1, points2, K1, K2) # pose.R, pose.t, pose.inliers Matching example in the usage guide →
Camera calibration
There is no checkerboard calibrator. Calibration models estimate intrinsics, gravity, roll, pitch and field of view from a single image; the stereo geometry guide reads calibration files that come from a calibration tool.