# -*- coding: utf-8 -*- """ Export COLMAP-estimated camera intrinsics + extrinsics to readable JSON/TXT, and validate against the provided ground-truth transforms.json / metadata.json. Input: COLMAP sparse model in TXT format (cameras.txt, images.txt) -> produced by `colmap model_converter --output_type TXT` Output: output/colmap/cameras_export/cameras.json (per-camera intrinsics) output/colmap/cameras_export/transforms.json (nerfstudio/3DGS-compatible c2w) prints: estimated intrinsics vs GT, trajectory alignment error (Umeyama) """ import os, sys, json import numpy as np def quat_to_rot(q): """COLMAP quaternion (w,x,y,z) -> rotation matrix (world->camera).""" w, x, y, z = q return np.array([ [1 - 2*(y*y + z*z), 2*(x*y - z*w), 2*(x*z + y*w)], [2*(x*y + z*w), 1 - 2*(x*x + z*z), 2*(y*z - x*w)], [2*(x*z - y*w), 2*(y*z + x*w), 1 - 2*(x*x + y*y)], ], dtype=np.float64) def parse_cameras_txt(path): cams = {} with open(path) as f: for line in f: line = line.strip() if not line or line.startswith('#'): continue t = line.split() cid = int(t[0]); model = t[1]; w = int(t[2]); h = int(t[3]) params = [float(x) for x in t[4:]] cams[cid] = dict(model=model, width=w, height=h, params=params) for cid, c in cams.items(): m = c['model'] if m in ('PINHOLE', 'OPENCV', 'SIMPLE_RADIAL', 'RADIAL'): fx, fy = c['params'][0], c['params'][1] cx, cy = c['params'][2], c['params'][3] if m in ('SIMPLE_RADIAL', 'RADIAL'): fx = fy = c['params'][0]; cx = c['params'][1]; cy = c['params'][2] elif m == 'SIMPLE_PINHOLE': fx = fy = c['params'][0]; cx = c['params'][1]; cy = c['params'][2] else: fx = fy = cx = cy = None c['fx'], c['fy'], c['cx'], c['cy'] = fx, fy, cx, cy return cams def parse_images_txt(path): imgs = [] with open(path) as f: lines = f.readlines() i = 0 while i < len(lines): line = lines[i].strip() if not line or line.startswith('#'): i += 1 continue t = line.split() iid = int(t[0]); q = np.array([float(x) for x in t[1:5]]) tvec = np.array([float(x) for x in t[5:8]]) cam_id = int(t[8]); name = t[9] i += 1 if i < len(lines): # skip the points2D line i += 1 R = quat_to_rot(q) # world -> camera c2w = np.eye(4) c2w[:3, :3] = R.T c2w[:3, 3] = -R.T @ tvec imgs.append(dict(id=iid, name=name, cam_id=cam_id, q=q, t=tvec, R_w2c=R, c2w=c2w)) return imgs def umeyama(src, dst): """Similarity transform aligning src->dst (both Nx3). Returns s,R,t,rms.""" src = np.asarray(src, np.float64); dst = np.asarray(dst, np.float64) mu_s = src.mean(0); mu_d = dst.mean(0) s = src - mu_s; d = dst - mu_d cov = d.T @ s / s.shape[0] U, S, Vt = np.linalg.svd(cov) R = U @ Vt if np.linalg.det(R) < 0: U[:, -1] *= -1; R = U @ Vt var_s = (s ** 2).sum() / s.shape[0] scale = np.trace(np.diag(S)) / var_s t = mu_d - scale * R @ mu_s aligned = (scale * R @ src.T).T + t rms = np.sqrt(((aligned - dst) ** 2).sum(1).mean()) return scale, R, t, rms, aligned def main(): in_dir = sys.argv[1] if len(sys.argv) > 1 else r"output\colmap\sparse\0" out_dir = sys.argv[2] if len(sys.argv) > 2 else r"output\colmap\cameras_export" os.makedirs(out_dir, exist_ok=True) cams = parse_cameras_txt(os.path.join(in_dir, "cameras.txt")) imgs = parse_images_txt(os.path.join(in_dir, "images.txt")) print(f"parsed {len(cams)} camera(s), {len(imgs)} image(s)") # sort images by id for stable ordering imgs.sort(key=lambda x: x['id']) # --- export cameras.json (intrinsics) --- cams_json = {} for cid, c in cams.items(): cams_json[str(cid)] = { "model": c['model'], "width": c['width'], "height": c['height'], "fx": c['fx'], "fy": c['fy'], "cx": c['cx'], "cy": c['cy'], "params": c['params'], } with open(os.path.join(out_dir, "cameras.json"), "w") as f: json.dump(cams_json, f, indent=2) # --- export transforms.json (nerfstudio/3DGS compatible, OpenCV convention) --- # use the dominant camera (single_camera=1 -> one camera) cam = cams[next(iter(cams))] fx, fy, cx, cy = cam['fx'], cam['fy'], cam['cx'], cam['cy'] w, h = cam['width'], cam['height'] angle_x = 2 * np.arctan(0.5 * w / fx) angle_y = 2 * np.arctan(0.5 * h / fy) transform = { "camera_model": "OPENCV", "fl_x": fx, "fl_y": fy, "cx": cx, "cy": cy, "w": w, "h": h, "camera_angle_x": float(angle_x), "camera_angle_y": float(angle_y), "k1": 0.0, "k2": 0.0, "p1": 0.0, "p2": 0.0, "frames": [ {"file_path": "images/" + img['name'], "transform_matrix": img['c2w'].tolist()} for img in imgs ], } with open(os.path.join(out_dir, "transforms.json"), "w") as f: json.dump(transform, f, indent=2) # --- also write a compact c2w numpy file for our own NeRF --- np.savez(os.path.join(out_dir, "poses_c2w.npz"), names=np.array([img['name'] for img in imgs]), c2w=np.stack([img['c2w'] for img in imgs]), fx=fx, fy=fy, cx=cx, cy=cy, w=w, h=h) # --- validation against GT --- print("\n===== VALIDATION =====") gt_meta = json.load(open(r"mvs_project\metadata.json")) gt_intr = gt_meta['intrinsics'] print(f"estimated intrinsics: fx={fx:.3f} fy={fy:.3f} cx={cx:.3f} cy={cy:.3f}") print(f"ground-truth intrinsics: fx={gt_intr['fx']:.3f} fy={gt_intr['fy']:.3f} " f"cx={gt_intr['cx']:.3f} cy={gt_intr['cy']:.3f}") print(f"intrinsics error: fx {abs(fx-gt_intr['fx']):.3f}px, fy {abs(fy-gt_intr['fy']):.3f}px, " f"cx {abs(cx-gt_intr['cx']):.3f}px, cy {abs(cy-gt_intr['cy']):.3f}px") # trajectory alignment: camera centers COLMAP vs GT gt_tr = json.load(open(r"mvs_project\poses\transforms.json")) by_name = {} for fr in gt_tr['frames']: by_name[fr['file_path'].split('/')[-1]] = np.array(fr['transform_matrix']) C_est, C_gt = [], [] common = [] for img in imgs: if img['name'] in by_name: common.append(img['name']) C_est.append(img['c2w'][:3, 3]) C_gt.append(by_name[img['name']][:3, 3]) C_est = np.array(C_est); C_gt = np.array(C_gt) s, R, t, rms, aligned = umeyama(C_est, C_gt) print(f"\ntrajectory: {len(common)} common views aligned (similarity s={s:.4f})") print(f"camera-center RMS after similarity alignment: {rms:.4f} m") print("(COLMAP recovers extrinsics up to a similarity gauge; " "this RMS measures trajectory-shape accuracy)") print("\nwrote:") print(" ", os.path.join(out_dir, "cameras.json")) print(" ", os.path.join(out_dir, "transforms.json")) print(" ", os.path.join(out_dir, "poses_c2w.npz")) if __name__ == "__main__": main()