shaniaguo/VisTa3D

This is the official dataset of VisTa3D.

Components

• data
  • depth (raw sensor depth maps; scale is 1000.0)
  • gt_depth (groundtruth depth maps; scale is 256.0)
  • image (RGB images)
  • pcd_mask (object masks)
  • intrinsics.npy (camera intrinsics)
• accelerometer.csv (timestamped accelerometer measurements)
• gyroscope.csv (timestamped gyroscope measurements)
• frames.csv (synchronization between timestamps, realworld time, and frames)
• CameraTrajectory.txt (camera extrinsics, see 'Usage' section for more details)
• *_pcd_registered.ply (dense, 3d scanned point cloud; already registered to current scene)

Usage

• Loading depth as meters. Depth scale for raw depth is 1000.0, for gt depth is 256.0.

def load_depth(path, depth_scale):
    d = np.array(Image.open(path)).astype(np.float32)
    d_m = d / float(depth_scale)
    return d_m

• Loading camera intrinsics.

def load_intrinsics(intr_path):
    return np.load(intr_path)

• Loading camera extrinsics from CameraTrajectory.txt. Convert each line into a 4 x 4 camera pose matrix (c2w).

def load_extrinsics(traj_path):
    mats = []
    for line in open(traj_path):
        if line.startswith("#") or not line.strip():
            continue
        vals = list(map(float, line.split()))
        if len(vals) != 8:
            continue
        _, tx, ty, tz, qx, qy, qz, qw = vals
        R_c2w = R.from_quat([qx, qy, qz, qw]).as_matrix()
        T = np.eye(4)
        T[:3, :3] = R_c2w
        T[:3, 3] = [tx, ty, tz]
        mats.append(T)
    return mats