prepare_scene_occ.py

"""
bash:

python prepare_scene_occ.py \
--replica_root ./Replica_SLAM \
--preprocessed_dir ./Replica_OCC/preprocessed \
--out_dir ./Replica_OCC/global_occ_package \
--scenes office0 \
--obs_stride_frame 1 \
--obs_stride_pix 1 \
--mask_dilate 0 \
--obs_max_frames -1 \
--max_depth 10.0
"""

import os
import json
import argparse
import pickle
import numpy as np
from PIL import Image
from tqdm import tqdm
from sklearn.neighbors import KDTree

try:
    from scipy.ndimage import binary_dilation
    HAS_SCIPY = True
except Exception:
    HAS_SCIPY = False


def load_cam_params(replica_root: str):
    cam_json = os.path.join(replica_root, "cam_params.json")
    with open(cam_json, "r") as f:
        js = json.load(f)
    cam = js.get("camera", js)
    fx, fy, cx, cy = float(cam["fx"]), float(cam["fy"]), float(cam["cx"]), float(cam["cy"])
    w, h = int(cam["w"]), int(cam["h"])
    scale = float(cam["scale"])
    K = np.eye(3, dtype=np.float32)
    K[0, 0] = fx
    K[1, 1] = fy
    K[0, 2] = cx
    K[1, 2] = cy
    return K, (w, h), scale


def load_traj(traj_path: str):
    traj = np.loadtxt(traj_path)
    assert traj.ndim == 2 and traj.shape[1] == 16
    Ts = traj.reshape(-1, 4, 4).astype(np.float32)  # cam->world
    return Ts


def load_preprocessed_scene(preprocessed_path: str):
    """
    Expect either (N,4) [x,y,z,label] or (N,7) [x,y,z,r,g,b,label].
    Return xyz (N,3), label (N,)
    """
    v = np.load(preprocessed_path)
    if v.ndim != 2 or v.shape[1] not in (4, 7):
        raise ValueError(f"Unexpected preprocessed shape: {v.shape}, path={preprocessed_path}")

    if v.shape[1] == 4:
        xyz = v[:, :3].astype(np.float32)
        lab = v[:, 3].astype(np.int32)
    else:
        xyz = v[:, :3].astype(np.float32)
        lab = v[:, 6].astype(np.int32)
    return xyz, lab


def build_regular_grid(xyz_shifted, voxel_size=0.08):
    """xyz_shifted already has min at ~0. Return grid_pts (Nx,Ny,Nz,3) and dims."""
    xyz_min = xyz_shifted.min(axis=0)
    xyz_max = xyz_shifted.max(axis=0)

    # snap to voxel grid
    xyz_min = np.floor(xyz_min / voxel_size) * voxel_size
    xyz_max = np.ceil(xyz_max / voxel_size) * voxel_size

    xs = np.arange(xyz_min[0], xyz_max[0] + 1e-9, voxel_size, dtype=np.float32)
    ys = np.arange(xyz_min[1], xyz_max[1] + 1e-9, voxel_size, dtype=np.float32)
    zs = np.arange(xyz_min[2], xyz_max[2] + 1e-9, voxel_size, dtype=np.float32)

    gx, gy, gz = np.meshgrid(xs, ys, zs, indexing="ij")
    grid_pts = np.stack([gx, gy, gz], axis=-1)  # (Nx,Ny,Nz,3)
    dims = [grid_pts.shape[0], grid_pts.shape[1], grid_pts.shape[2]]
    return grid_pts, dims


def assign_labels_to_grid(grid_pts, xyz, lab, voxel_size=0.08):
    """Nearest neighbor label assignment from sparse voxels to regular grid centers."""
    flat = grid_pts.reshape(-1, 3)
    tree = KDTree(xyz, leaf_size=32)
    dist, ind = tree.query(flat, k=1)
    dist = dist.reshape(-1)
    ind = ind.reshape(-1)

    out = np.zeros((flat.shape[0],), dtype=np.int32)
    m = dist <= voxel_size
    out[m] = lab[ind[m]]
    return out.reshape(grid_pts.shape[:3])


def build_scene_mask_by_fused_frustums(
    replica_root,
    scene,
    grid_pts,          # (Nx,Ny,Nz,3) float **in world coordinates**
    Ts_cw,             # (F,4,4) cam->world
    K,                 # (3,3)
    depth_scale,
    max_frames=-1,
    stride_frame=5,
    obs_stride_pix=1,  # NOTE: currently not used (voxel-domain projection); keep for API compatibility
    max_depth=10.0,
    tau=0.08,          # depth tolerance ~ 1 voxel
):
    """Scene-level mask via per-frame frustum + depth test (union across frames).

    grid_pts is already in the Replica world coordinate system after adding origin_shift back.
    No extra translation is applied inside this function, keeping it consistent with camera poses.
    """
    scene_dir = os.path.join(replica_root, scene)
    depth_dir = os.path.join(scene_dir, "depths")
    if not os.path.isdir(depth_dir):
        return np.zeros(grid_pts.shape[:3], dtype=np.float32), []

    depth_files = sorted([f for f in os.listdir(depth_dir) if f.endswith(".png")])
    if max_frames > 0:
        depth_files = depth_files[:max_frames]

    Nx, Ny, Nz = grid_pts.shape[:3]
    mask = np.zeros((Nx, Ny, Nz), dtype=np.uint8)
    used_imgs = []

    # grid points are already in world coordinates
    Pw = grid_pts.reshape(-1, 3).astype(np.float32)

    fx, fy, cx, cy = K[0, 0], K[1, 1], K[0, 2], K[1, 2]

    it = enumerate(depth_files)
    it = tqdm(list(it), desc=f"[mask] {scene}", leave=False) if len(depth_files) > 50 else it

    for _, fname in it:
        # stride over frames
        # fname like depth000123.png
        idx = int(fname.replace("depth", "").replace(".png", ""))
        if (idx % stride_frame) != 0:
            continue
        if idx >= Ts_cw.shape[0]:
            continue

        depth_png = os.path.join(depth_dir, fname)
        d16 = np.array(Image.open(depth_png).convert("I;16"), dtype=np.float32)
        depth_m = d16 / float(depth_scale)
        depth_m[depth_m > max_depth] = 0.0
        H, W = depth_m.shape

        # cam->world
        Twc = Ts_cw[idx]
        Rwc = Twc[:3, :3]
        twc = Twc[:3, 3]

        # world -> cam: Pc = Rcw*(Pw - twc), where Rcw = Rwc^T
        Rcw = Rwc.T
        Pc = (Rcw @ (Pw - twc[None, :]).T).T  # (N,3)

        z_all = Pc[:, 2]

        # 0) finite + in_front
        finite = np.isfinite(Pc).all(axis=1) & np.isfinite(z_all)
        in_front = z_all > 1e-6
        ok = finite & in_front
        if not np.any(ok):
            continue

        idx_ok = np.nonzero(ok)[0]   # keep original Pw indices
        Pc_ok = Pc[idx_ok]

        x = Pc_ok[:, 0]
        y = Pc_ok[:, 1]
        z = Pc_ok[:, 2]

        # 1) project
        u = fx * (x / z) + cx
        v = fy * (y / z) + cy

        # 2) u,v finite
        uv_ok = np.isfinite(u) & np.isfinite(v)
        if not np.any(uv_ok):
            continue

        idx_uv = idx_ok[uv_ok]
        u = u[uv_ok]
        v = v[uv_ok]
        z = z[uv_ok]

        ui = np.rint(u).astype(np.int32)
        vi = np.rint(v).astype(np.int32)

        # 3) in image
        in_img = (ui >= 0) & (ui < W) & (vi >= 0) & (vi < H)
        if not np.any(in_img):
            continue

        idx_img = idx_uv[in_img]
        ui = ui[in_img]
        vi = vi[in_img]
        z = z[in_img]

        # 4) depth lookup + validity
        di = depth_m[vi, ui]
        good_depth = di > 1e-6
        if not np.any(good_depth):
            continue

        idx_depth = idx_img[good_depth]
        zi = z[good_depth]
        di = di[good_depth]

        valid = zi <= (di + tau)
        idxs = idx_depth[valid]  # FINAL: indices into Pw

        if idxs.size > 0:
            m = np.zeros((Pw.shape[0],), dtype=np.uint8)
            m[idxs] = 1
            mask |= m.reshape(Nx, Ny, Nz)

        rgb_path = os.path.join(scene_dir, "frames", f"frame{idx:06d}.jpg")
        used_imgs.append(rgb_path)

    return mask.astype(np.float32), used_imgs


def optional_mask_dilation(mask, dilate_iter=0):
    if dilate_iter <= 0:
        return mask
    if not HAS_SCIPY:
        print("[Warn] scipy not installed; skip dilation.")
        return mask
    return binary_dilation(mask.astype(np.uint8), iterations=dilate_iter).astype(np.float32)


def process_one_scene(
    replica_root,
    out_root,
    scene,
    preprocessed_dir,
    voxel_size=0.08,
    max_depth=10.0,
    obs_max_frames=-1,
    obs_stride_frame=5,
    obs_stride_pix=1,
    mask_dilate=0,
):
    pre_path = os.path.join(preprocessed_dir, f"{scene}.npy")
    xyz, lab = load_preprocessed_scene(pre_path)

    # Minimum point in the original Replica world coordinates, used to build a min-aligned regular grid.
    origin_shift = xyz.min(axis=0)
    xyz_shifted = xyz - origin_shift[None, :]

    # 1) Build a regular grid in the shifted coordinate system.
    grid_pts_shifted, dims = build_regular_grid(xyz_shifted, voxel_size=voxel_size)

    # 2) Assign labels by nearest-neighbor lookup in the same shifted coordinate system.
    global_labels = assign_labels_to_grid(grid_pts_shifted, xyz_shifted, lab, voxel_size=voxel_size)

    # 3) Translate grid points back to the Replica world coordinate system.
    grid_pts_world = grid_pts_shifted + origin_shift[None, None, None, :]

    # 4) Build the observed-space mask in world coordinates using camera frustums and depth maps.
    K, (_W, _H), depth_scale = load_cam_params(replica_root)
    Ts = load_traj(os.path.join(replica_root, scene, "traj.txt"))  # cam->world

    global_mask, used_imgs = build_scene_mask_by_fused_frustums(
        replica_root=replica_root,
        scene=scene,
        grid_pts=grid_pts_world,   # Pass world-coordinate grid points directly.
        Ts_cw=Ts,
        K=K,
        depth_scale=depth_scale,
        max_frames=obs_max_frames,
        stride_frame=obs_stride_frame,
        obs_stride_pix=obs_stride_pix,
        max_depth=max_depth,
        tau=voxel_size,
    )

    global_mask = optional_mask_dilation(global_mask, dilate_iter=mask_dilate)

    # mask=0 -> unknown
    unknown = (global_mask < 0.5)
    global_labels[unknown] = 255

    # mask=1 and label=0 -> known-free remains 0 for completion evaluation.
    known_free = (global_mask >= 0.5) & (global_labels == 0)
    global_labels[known_free] = 0

    out = {
        "scene_name": scene,
        "scene_dim": dims,
        "global_labels": global_labels.astype(np.int64),
        # global_pts saved in the pkl are voxel centers in Replica world coordinates.
        "global_pts": grid_pts_world.astype(np.float64),
        "valid_img_count": len(used_imgs),
        "valid_img_paths": used_imgs,
        "global_mask": global_mask.astype(np.float64),
    }

    os.makedirs(out_root, exist_ok=True)
    save_path = os.path.join(out_root, f"{scene}.pkl")
    with open(save_path, "wb") as f:
        pickle.dump(out, f, protocol=pickle.HIGHEST_PROTOCOL)

    print(
        f"[OK] {scene}: dim={dims}, labels={np.unique(global_labels).size}, "
        f"mask_mean={global_mask.mean():.3f}, saved={save_path}"
    )


def main():
    ap = argparse.ArgumentParser()
    ap.add_argument("--replica_root", type=str, required=True,
                    help="Replica_SLAM root (has cam_params.json + scene folders)")
    ap.add_argument("--preprocessed_dir", type=str, required=True,
                    help="global preprocessed dir (preprocessed/*.npy)")
    ap.add_argument("--out_dir", type=str, required=True,
                    help="output dir for scene-level pkls")
    ap.add_argument("--scenes", type=str, default="",
                    help="comma-separated scenes; empty means all under replica_root")
    ap.add_argument("--voxel_size", type=float, default=0.08)
    ap.add_argument("--max_depth", type=float, default=10.0)

    ap.add_argument("--obs_max_frames", type=int, default=-1)
    ap.add_argument("--obs_stride_frame", type=int, default=5)
    ap.add_argument("--obs_stride_pix", type=int, default=1)
    ap.add_argument("--mask_dilate", type=int, default=0)

    args = ap.parse_args()

    if args.scenes.strip():
        scene_list = [s.strip() for s in args.scenes.split(",") if s.strip()]
    else:
        scene_list = []
        for s in sorted(os.listdir(args.replica_root)):
            p = os.path.join(args.replica_root, s, "traj.txt")
            if os.path.isfile(p):
                scene_list.append(s)

    for scene in scene_list:
        process_one_scene(
            replica_root=args.replica_root,
            out_root=args.out_dir,
            scene=scene,
            preprocessed_dir=args.preprocessed_dir,
            voxel_size=args.voxel_size,
            max_depth=args.max_depth,
            obs_max_frames=args.obs_max_frames,
            obs_stride_frame=args.obs_stride_frame,
            obs_stride_pix=args.obs_stride_pix,
            mask_dilate=args.mask_dilate,
        )


if __name__ == "__main__":
    main()