import sys, json, time, math import numpy as np, onnxruntime as ort from PIL import Image from scipy.optimize import least_squares from functools import partial def uv_grid(W, H): ar = W / H sx = ar / math.sqrt(1 + ar**2); sy = 1 / math.sqrt(1 + ar**2) u = np.linspace(-sx * (W - 1) / W, sx * (W - 1) / W, W) v = np.linspace(-sy * (H - 1) / H, sy * (H - 1) / H, H) u, v = np.meshgrid(u, v, indexing='xy') return np.stack([u, v], -1) def nearest_idx(n_in, n_out): # torch F.interpolate mode='nearest': src = floor(dst * in/out) return np.minimum(np.floor(np.arange(n_out) * (n_in / n_out)).astype(int), n_in - 1) def solve_focal_shift(uv, xyz): uv, xy, z = uv.reshape(-1, 2), xyz[..., :2].reshape(-1, 2), xyz[..., 2].reshape(-1) def fn(shift): xy_proj = xy / (z + shift)[:, None] f = (xy_proj * uv).sum() / np.square(xy_proj).sum() return (f * xy_proj - uv).ravel() sol = least_squares(fn, x0=0, ftol=1e-3, method='lm') s = float(sol['x'].squeeze()) xy_proj = xy / (z + s)[:, None] f = (xy_proj * uv).sum() / np.square(xy_proj).sum() return s, float(f) def postprocess(points, mask_p, scale): H, W = mask_p.shape mask = mask_p > 0.5 uv = uv_grid(W, H) ri, ci = nearest_idx(H, 64), nearest_idx(W, 64) p_lr = points[ri][:, ci]; uv_lr = uv[ri][:, ci]; m_lr = mask[ri][:, ci] shift, focal = solve_focal_shift(uv_lr[m_lr], p_lr[m_lr]) ar = W / H fx = focal / 2 * math.sqrt(1 + ar**2) / ar; fy = focal / 2 * math.sqrt(1 + ar**2) depth = points[..., 2] + shift mask &= depth > 0 u = (np.arange(W) + 0.5) / W; v = (np.arange(H) + 0.5) / H uu, vv = np.meshgrid(u, v) X = (uu - 0.5) / fx * depth; Y = (vv - 0.5) / fy * depth P = np.stack([X, Y, depth], -1) * scale return dict(points=P, depth=depth * scale, mask=mask, fx=fx, fy=fy, shift=shift, focal=focal, fov_x=math.degrees(2 * math.atan(0.5 / fx)), fov_y=math.degrees(2 * math.atan(0.5 / fy))) def run(path, long_side=512, num_tokens=1800, sess=None): im = Image.open(path).convert('RGB') w, h = im.size; s = long_side / max(w, h) W, H = round(w * s), round(h * s) im = im.resize((W, H), Image.BICUBIC) x = (np.asarray(im).astype(np.float32) / 255).transpose(2, 0, 1)[None] t = time.time() out = sess.run(None, {'image': x, 'num_tokens': np.array(num_tokens, dtype=np.int64)}) dt = time.time() - t pts, nrm, msk, sc = out r = postprocess(pts[0].astype(np.float64), msk[0], float(sc[0])) r.update(ms=dt * 1000, W=W, H=H, scale=float(sc[0]), normal=nrm[0]) return r if __name__ == '__main__': so = ort.SessionOptions(); so.intra_op_num_threads = 4 sess = ort.InferenceSession('model.onnx', so, providers=['CPUExecutionProvider']) res = {} for f in sys.argv[1:]: r = run(f, sess=sess) d = r['depth'][r['mask']] info = {k: r[k] for k in ['W', 'H', 'ms', 'scale', 'shift', 'focal', 'fov_x', 'fov_y']} info.update(depth_min=float(d.min()), depth_med=float(np.median(d)), depth_max=float(d.max()), mask_pct=float(r['mask'].mean() * 100)) print(f, json.dumps(info)) import os; np.save('/workspace/geometria3d-lab/_work/' + os.path.basename(f) + '.points.npy', r['points'].astype(np.float32)) res[f] = info