Introduction

• Two‑view SfM: ORB features → Essential matrix (RANSAC) → pose recovery → linear triangulation → colored point cloud (PLY).
• PnP: Camera re‑localization from 2D–3D matches.
• RGB‑D pose+depth fusion: fuse depth frames using known/external poses into a single point cloud.
• Synthetic pair generator: sanity‑check geometry with clean correspondences.
• Tiny bundle‑adjustment (BA) with scipy.optimize.least_squares (optional, small problems).

Focus: clarity over speed. Few dependencies, runs locally.

Install

pip install -r requirements.txt

Quickstart

1) Two‑view SfM on your own images

Put two images in data/sample_pair/ as im1.jpg, im2.jpg (roughly same scene).

python scripts/two_view_sfm.py --im1 data/sample_pair/im1.jpg --im2 data/sample_pair/im2.jpg     --fx 1200 --fy 1200 --cx 960 --cy 540 --out cloud.ply

2) Synthetic correspondences sanity‑check

python scripts/synth_pair.py --n_points 300 --noise_px 0.5

Prints recovered pose, saves a small PLY (synth_cloud.ply).

3) Fuse depth + poses into one cloud (RGB‑D or monocular+depth)

Assumes per‑frame depth_*.npy (meters), per‑frame pose_*.txt (4×4) and intrinsics K.json:

python scripts/fuse_depth_pose.py --depth_dir data/rgbd/depth --pose_dir data/rgbd/poses     --K data/rgbd/K.json --stride 1 --out fused_cloud.ply

Notes

• Intrinsics: pass as --fx --fy --cx --cy. If unknown, try EXIF focal length + sensor width or calibrate with a checkerboard.
• PLY is ASCII for easy inspection (CloudCompare, MeshLab).
• For real videos → run feature tracking outside the scope here (e.g., KLT) or extend this repo.