"""Export metric DAV2 Small depth for every video frame.""" import argparse import csv import json from pathlib import Path import cv2 import numpy as np import torch from huggingface_hub import hf_hub_download from depth_anything_v2.dpt import DepthAnythingV2 from geometry import slope_features MODEL = 'depth-anything/Depth-Anything-V2-Metric-VKITTI-Small' WEIGHTS = 'depth_anything_v2_metric_vkitti_vits.pth' def main(): parser = argparse.ArgumentParser(description=__doc__) parser.add_argument('--video', type=Path, required=True) parser.add_argument('--frames-csv', type=Path) parser.add_argument('--geometry-csv', type=Path, help='Per-frame intrinsics and optional camera up vectors') parser.add_argument('--out', type=Path, required=True) parser.add_argument('--checkpoint', type=Path, help='Use local weights instead of downloading') parser.add_argument('--limit', type=int, help='Only process the first N frames for a trial') args = parser.parse_args() if args.limit is not None and args.limit <= 0: parser.error('--limit must be positive') if args.out.exists() and any(args.out.iterdir()): parser.error('--out must be empty to avoid mixing clips') capture = cv2.VideoCapture(str(args.video)) count = int(capture.get(cv2.CAP_PROP_FRAME_COUNT)) fps = capture.get(cv2.CAP_PROP_FPS) width = int(capture.get(cv2.CAP_PROP_FRAME_WIDTH)) height = int(capture.get(cv2.CAP_PROP_FRAME_HEIGHT)) if not capture.isOpened() or count <= 0 or fps <= 0: raise ValueError(f'Cannot read video: {args.video}') source = None if args.frames_csv: with args.frames_csv.open(encoding='utf-8-sig', newline='') as stream: source = list(csv.DictReader(stream)) if [int(row['frame_idx']) for row in source] != list(range(count)): raise ValueError('CSV frame_idx must match every video frame in order') geometry = None gravity_aligned = False if args.geometry_csv: with args.geometry_csv.open(encoding='utf-8-sig', newline='') as stream: geometry = list(csv.DictReader(stream)) if [int(row['frame_idx']) for row in geometry] != list(range(count)): raise ValueError('Geometry CSV must match every video frame in order') gravity_aligned = all(key in geometry[0] for key in ['up_x','up_y','up_z']) for index, params in enumerate(geometry): if (int(params['width']), int(params['height'])) != (width, height): raise ValueError('Intrinsics resolution must match the video; use native fields for native video') if source is not None and 'stamp' in params and abs(float(params['stamp'])-float(source[index]['stamp'])) > 1e-5: raise ValueError(f'Geometry/frame timestamps differ at frame {index}') torch.set_num_threads(4) device = 'cuda' if torch.cuda.is_available() else 'cpu' checkpoint = str(args.checkpoint) if args.checkpoint else hf_hub_download(MODEL, WEIGHTS) model = DepthAnythingV2(encoder='vits', features=64, out_channels=[48, 96, 192, 384], max_depth=80) model.load_state_dict(torch.load(checkpoint, map_location='cpu', weights_only=True)) model = model.to(device).eval() (args.out / 'depth').mkdir(parents=True) if geometry is not None: (args.out / 'terrain').mkdir() # Compact previews; raw arrays retain the source resolution. pw = min(width, 960) // 2 * 2 ph = round(height * pw / width) // 2 * 2 writers = [cv2.VideoWriter(str(args.out / name), cv2.VideoWriter_fourcc(*'mp4v'), fps, size) for name, size in [ ('depth.mp4', (pw, ph)), ('preview.mp4', (pw * (3 if geometry is not None else 2), ph))]] rows = [] try: if not all(writer.isOpened() for writer in writers): raise RuntimeError('MP4 encoder could not be opened') total = min(count, args.limit) if args.limit else count with torch.inference_mode(): for index in range(total): ok, image = capture.read() if not ok: raise RuntimeError(f'Cannot decode frame {index}') depth = model.infer_image(image, 518).astype(np.float32) if depth.shape != (height, width) or not np.isfinite(depth).all(): raise RuntimeError(f'Invalid depth at frame {index}') name = f'depth/frame{index:06d}.npy' np.save(args.out / name, depth) colored = cv2.applyColorMap(np.uint8(np.clip(depth / 50, 0, 1) * 255), cv2.COLORMAP_TURBO) colored = cv2.resize(colored, (pw, ph)) writers[0].write(colored) panels = [cv2.resize(image, (pw, ph)), colored] if geometry is not None: params = geometry[index] fx,fy,cx,cy = [float(params[key]) for key in ['fx','fy','cx','cy']] k = np.array([[fx,0,cx],[0,fy,cy],[0,0,1]], dtype=np.float64) up = np.array([float(params[key]) for key in ['up_x','up_y','up_z']]) if gravity_aligned else np.array([0.,-1.,0.]) if not np.isfinite(k).all() or fx <= 0 or fy <= 0 or not np.isfinite(up).all() or np.linalg.norm(up) == 0: raise ValueError(f'Invalid geometry at frame {index}') slope, valid = slope_features(depth, k, up) field = 'slope_degrees' if gravity_aligned else 'slope_degrees_proxy' np.savez_compressed(args.out / 'terrain' / f'frame{index:06d}.npz', **{field: slope.astype(np.float16), 'valid': valid}) pixels = np.uint8(np.clip(np.nan_to_num(slope) / 90, 0, 1) * 255) panel = cv2.applyColorMap(pixels, cv2.COLORMAP_TURBO) panel[~valid] = 0 panel = cv2.resize(panel, (pw, ph)) cv2.putText(panel, 'Slope estimate 0-90 deg' if gravity_aligned else 'Slope proxy 0-90 deg', (8,24), cv2.FONT_HERSHEY_SIMPLEX, .6, (255,255,255), 1) panels.append(panel) writers[1].write(np.hstack(panels)) row = dict(source[index]) if source is not None else dict(frame_idx=index) row.update(source_frame=index, timestamp_s=float(row['stamp']) if source else index / fps, depth=name, height=height, width=width, depth_units='meters') if geometry is not None: row.update(terrain=f'terrain/frame{index:06d}.npz', geometry_reference='recorded_intrinsics_arkit_gravity' if gravity_aligned else 'recorded_intrinsics_assumed_up') rows.append(row) print(f'{index + 1}/{total}', flush=True) finally: capture.release() for writer in writers: writer.release() with (args.out / 'frames.csv').open('w', encoding='utf-8', newline='') as stream: writer = csv.DictWriter(stream, fieldnames=list(rows[0])) writer.writeheader() writer.writerows(rows) report = dict(model=MODEL, device=device, source_frames=count, exported_frames=len(rows), source_fps=fps, depth_shape=[height, width], depth_dtype='float32', depth_units='meters (model prediction)', input_size=518, preview_color_range_m=[0, 50], preview_layout='RGB | depth | slope' if geometry is not None else 'RGB | depth', recorded_intrinsics_used=geometry is not None, gravity_aligned=gravity_aligned, depth_scale_validated=False, timestamp_source='CSV stamp' if source else 'frame index / video FPS') (args.out / 'run.json').write_text(json.dumps(report, indent=2), encoding='utf-8') if __name__ == '__main__': main()