vis_scene_occ.py

"""
bash:

python vis_scene_occ.py   --pkl ./Replica_OCC/global_occ_package/office0.pkl   --downsample 1
"""

import argparse
import pickle
import numpy as np
import open3d as o3d


def label_to_color(labels: np.ndarray) -> np.ndarray:
    """
    Deterministic pseudo-color for semantic labels (>0).
    labels: (N,)
    returns: (N,3) in [0,1]
    """
    labels = labels.astype(np.int64)
    r = ((labels * 37) % 255) / 255.0
    g = ((labels * 17) % 255) / 255.0
    b = ((labels * 97) % 255) / 255.0
    return np.stack([r, g, b], axis=1).astype(np.float32)


def _as_numpy(x):
    if isinstance(x, np.ndarray):
        return x
    return np.asarray(x)


def resize_mask_to_labels_nn(mask_3d: np.ndarray, target_shape):
    """
    Nearest-neighbor resize for 3D mask.
    mask_3d: (sx,sy,sz)
    target_shape: (tx,ty,tz)
    return: (tx,ty,tz) float32
    """
    src = _as_numpy(mask_3d)
    if src.ndim != 3:
        raise ValueError(f"resize_mask_to_labels_nn expects 3D, got {src.ndim}D shape={src.shape}")

    if tuple(src.shape) == tuple(target_shape):
        return src.astype(np.float32)

    sx, sy, sz = src.shape
    tx, ty, tz = target_shape

    # nearest-neighbor index mapping
    xs = np.round(np.linspace(0, sx - 1, tx)).astype(np.int64)
    ys = np.round(np.linspace(0, sy - 1, ty)).astype(np.int64)
    zs = np.round(np.linspace(0, sz - 1, tz)).astype(np.int64)

    out = src[np.ix_(xs, ys, zs)]
    return out.astype(np.float32)


def detect_format(d: dict):
    """
    Return mode: 'dense' or 'sparse'
    Also return extracted (pts, labels, mask) in some raw form.
    Accepts:
      Dense:
        global_pts: (Nx,Ny,Nz,3)
        global_labels: (Nx,Ny,Nz) or (Nx,Ny,Nz,1)
        global_mask: (Nx,Ny,Nz) OR other shape (will be resized)
      Sparse:
        global_pts: (N,3)
        global_labels: (N,)
        global_mask: (N,)
    """
    if "global_pts" not in d:
        raise KeyError("pkl missing key: global_pts")

    pts = _as_numpy(d["global_pts"])

    # labels key name could vary; try common ones
    if "global_labels" in d:
        labels = _as_numpy(d["global_labels"])
    elif "labels" in d:
        labels = _as_numpy(d["labels"])
    else:
        labels = None

    # mask key name could vary; try common ones
    if "global_mask" in d:
        mask = _as_numpy(d["global_mask"])
    elif "mask" in d:
        mask = _as_numpy(d["mask"])
    else:
        mask = None

    # Dense if pts is 4D (...,3)
    if pts.ndim == 4 and pts.shape[-1] == 3:
        return "dense", pts, labels, mask

    # Sparse if pts is 2D (N,3)
    if pts.ndim == 2 and pts.shape[1] == 3:
        return "sparse", pts, labels, mask

    # Fallback: try to interpret any (N,3) shaped last dim
    if pts.shape[-1] == 3:
        return "sparse", pts.reshape(-1, 3), labels, mask

    raise ValueError(f"Unrecognized global_pts shape: {pts.shape}")


def coerce_dense_labels(labels: np.ndarray, target_shape):
    """
    Make labels dense grid of shape target_shape (Nx,Ny,Nz).
    labels can be (Nx,Ny,Nz), (Nx,Ny,Nz,1), or flat (N,).
    """
    if labels is None:
        return np.zeros(target_shape, dtype=np.int32)

    lab = _as_numpy(labels)

    if lab.ndim == 3 and tuple(lab.shape) == tuple(target_shape):
        return lab.astype(np.int32)

    if lab.ndim == 4 and lab.shape[-1] == 1 and tuple(lab.shape[:3]) == tuple(target_shape):
        return lab[..., 0].astype(np.int32)

    # flat
    if lab.ndim == 1:
        if lab.size == int(np.prod(target_shape)):
            return lab.reshape(target_shape).astype(np.int32)

    # last resort: try reshape if total matches
    if lab.size == int(np.prod(target_shape)):
        return lab.reshape(target_shape).astype(np.int32)

    print(f"[Warn] Cannot coerce labels to target shape. labels shape={lab.shape}, target={target_shape}. Use zeros.")
    return np.zeros(target_shape, dtype=np.int32)


def coerce_dense_mask(mask: np.ndarray, target_shape, allow_resize=True):
    """
    Make mask dense grid of shape target_shape (Nx,Ny,Nz).
    mask can be:
      - (Nx,Ny,Nz)
      - different 3D (will NN-resize if allow_resize)
      - flat (N,)
    """
    if mask is None:
        return np.ones(target_shape, dtype=np.float32)

    m = _as_numpy(mask)

    # already matching
    if m.ndim == 3 and tuple(m.shape) == tuple(target_shape):
        return m.astype(np.float32)

    # different 3D -> resize
    if m.ndim == 3 and allow_resize:
        print(f"[Warn] mask shape {m.shape} != labels shape {target_shape}, resizing mask by NN")
        return resize_mask_to_labels_nn(m, target_shape)

    # flat -> reshape if possible
    if m.ndim == 1:
        if m.size == int(np.prod(target_shape)):
            return m.reshape(target_shape).astype(np.float32)

    # last resort: if sizes match, reshape anyway
    if m.size == int(np.prod(target_shape)):
        return m.reshape(target_shape).astype(np.float32)

    print(f"[Warn] Cannot coerce mask to target shape. mask shape={m.shape}, target={target_shape}. Use ones.")
    return np.ones(target_shape, dtype=np.float32)


def coerce_sparse_labels(labels: np.ndarray, n_points: int):
    if labels is None:
        return np.zeros((n_points,), dtype=np.int32)
    lab = _as_numpy(labels).reshape(-1)
    if lab.size == n_points:
        return lab.astype(np.int32)
    print(f"[Warn] Sparse labels size mismatch: labels={lab.size}, pts={n_points}. Truncate/min-align.")
    n = min(lab.size, n_points)
    out = np.zeros((n_points,), dtype=np.int32)
    out[:n] = lab[:n].astype(np.int32)
    return out


def coerce_sparse_mask(mask: np.ndarray, n_points: int):
    if mask is None:
        return np.ones((n_points,), dtype=np.float32)
    m = _as_numpy(mask).reshape(-1)
    if m.size == n_points:
        return m.astype(np.float32)
    print(f"[Warn] Sparse mask size mismatch: mask={m.size}, pts={n_points}. Truncate/min-align.")
    n = min(m.size, n_points)
    out = np.ones((n_points,), dtype=np.float32)
    out[:n] = m[:n].astype(np.float32)
    return out


def main():
    parser = argparse.ArgumentParser()
    parser.add_argument("--pkl", type=str, required=True, help="scene-level occ pkl")
    parser.add_argument("--downsample", type=int, default=1,
                        help="visualization downsample stride (>=1), 1 = no downsample")
    args = parser.parse_args()

    with open(args.pkl, "rb") as f:
        d = pickle.load(f)

    mode, pts_raw, labels_raw, mask_raw = detect_format(d)

    # Print raw shapes
    try:
        print("[Info] global_pts shape:", _as_numpy(pts_raw).shape)
    except Exception:
        pass
    if labels_raw is not None:
        print("[Info] global_labels shape:", _as_numpy(labels_raw).shape)
    if mask_raw is not None:
        print("[Info] global_mask shape:", _as_numpy(mask_raw).shape)

    if mode == "dense":
        # pts: (Nx,Ny,Nz,3)
        pts_grid = _as_numpy(pts_raw).astype(np.float32)
        Nx, Ny, Nz = pts_grid.shape[:3]
        target_shape = (Nx, Ny, Nz)
        print("[Mode] Dense grid detected")
        print(f"[Info] scene_dim = {target_shape}")

        labels_grid = coerce_dense_labels(labels_raw, target_shape)
        mask_grid = coerce_dense_mask(mask_raw, target_shape, allow_resize=True)

        # flatten all aligned arrays
        pts = pts_grid.reshape(-1, 3)
        labels = labels_grid.reshape(-1)
        mask = mask_grid.reshape(-1)

    else:
        # sparse
        print("[Mode] Sparse points detected")
        pts = _as_numpy(pts_raw).reshape(-1, 3).astype(np.float32)
        labels = coerce_sparse_labels(labels_raw, pts.shape[0])
        mask = coerce_sparse_mask(mask_raw, pts.shape[0])

        print(f"[Info] points = {pts.shape[0]}")

    # optional downsample for speed (stride over flattened list)
    if args.downsample > 1 and pts.shape[0] > 0:
        sel = np.arange(0, pts.shape[0], args.downsample, dtype=np.int64)
        pts = pts[sel]
        labels = labels[sel]
        mask = mask[sel]

    # colors
    colors = np.zeros((pts.shape[0], 3), dtype=np.float32)

    # --- color rules ---
    # mask == 0  → dark gray
    mask0 = mask < 0.5
    colors[mask0] = np.array([0.9, 0.9, 0.9], dtype=np.float32)  # light gray

    # mask == 1 & label == 0 → light gray (empty but valid)
    mask1_empty = (mask >= 0.5) & (labels == 0)
    colors[mask1_empty] = np.array([0.2, 0.2, 0.2], dtype=np.float32)  # dark gray

    # mask == 1 & label > 0 → semantic color
    mask1_sem = (mask >= 0.5) & (labels > 0)
    if np.any(mask1_sem):
        colors[mask1_sem] = label_to_color(labels[mask1_sem])

    # build Open3D point cloud
    pcd = o3d.geometry.PointCloud()
    pcd.points = o3d.utility.Vector3dVector(pts.astype(np.float32))
    pcd.colors = o3d.utility.Vector3dVector(colors.astype(np.float32))

    # stats
    print("[Stats]")
    print("  total points:", pts.shape[0])
    print("  mask=1 ratio:", float((mask >= 0.5).mean()) if pts.shape[0] > 0 else 0.0)
    print("  semantic voxels:", int((labels > 0).sum()))

    o3d.visualization.draw_geometries(
        [pcd],
        window_name="Scene-level OCC (mask + semantic)",
        width=1280,
        height=960,
    )


if __name__ == "__main__":
    main()