prepare_preprocessed.py

"""
bash:

python prepare_preprocessed.py \
--replica_root ./Replica_SLAM \
--out_root ./Replica_OCC \
--stride 4 \
--depth_scale -1 \
--max_depth 10.0 \
--max_frames -1
"""

import os
import glob
import json
import argparse
from tqdm import tqdm
import numpy as np
import cv2
from PIL import Image
from mmengine import track_parallel_progress

# ---- globals for multiprocessing workers ----
_G_REPLICA_ROOT = None
_G_OUT_PREPROCESSED = None
_G_MAX_FRAMES = None
_G_STRIDE = None
_G_DEPTH_SCALE = None
_G_MAX_DEPTH = None


# -----------------------------
# Replica-specific IO
# -----------------------------

def _try_load_intrinsics(scene_dir):
    """Load intrinsics for Replica-SLAM format.
    Priority:
      1) scene_dir/* (intrinsic.txt, intrinsics.txt, cam.txt, camera.txt)
      2) dataset_root/cam_params.json  (sibling of scene folders)
    Returns:
      K4: 4x4 float32
      depth_scale_from_file: float or None
    """
    # 1) old candidates inside scene_dir
    candidates = [
        os.path.join(scene_dir, "intrinsic.txt"),
        os.path.join(scene_dir, "intrinsics.txt"),
        os.path.join(scene_dir, "cam.txt"),
        os.path.join(scene_dir, "camera.txt"),
    ]
    for p in candidates:
        if os.path.exists(p):
            K = np.loadtxt(p)
            K = np.asarray(K)
            if K.ndim == 1 and K.size == 4:
                fx, fy, cx, cy = K.tolist()
                KK = np.eye(4, dtype=np.float32)
                KK[0, 0] = fx
                KK[1, 1] = fy
                KK[0, 2] = cx
                KK[1, 2] = cy
                return KK, None
            if K.shape == (3, 3):
                KK = np.eye(4, dtype=np.float32)
                KK[:3, :3] = K
                return KK, None
            if K.shape == (4, 4):
                return K.astype(np.float32), None

    # 2) Replica-SLAM common: cam_params.json at dataset root (scene_dir/..)
    dataset_root = os.path.dirname(scene_dir)
    json_path = os.path.join(dataset_root, "cam_params.json")
    if os.path.exists(json_path):
        with open(json_path, "r") as f:
            js = json.load(f)
        cam = js.get("camera", js)  # tolerate either {"camera":{...}} or flat
        fx = float(cam["fx"])
        fy = float(cam["fy"])
        cx = float(cam["cx"])
        cy = float(cam["cy"])
        KK = np.eye(4, dtype=np.float32)
        KK[0, 0] = fx
        KK[1, 1] = fy
        KK[0, 2] = cx
        KK[1, 2] = cy

        depth_scale = cam.get("scale", None)
        depth_scale = float(depth_scale) if depth_scale is not None else None
        return KK, depth_scale

    raise FileNotFoundError(
        f"Cannot find intrinsics. Tried {candidates} and {json_path}"
    )


def _load_traj_as_Tcw_list(traj_path):
    """Load traj.txt and return list of 4x4 matrices extCam2World (Tcw)."""
    arr = np.loadtxt(traj_path)
    arr = np.asarray(arr)
    if arr.ndim == 3 and arr.shape[-2:] == (4, 4):
        return [arr[i] for i in range(arr.shape[0])]
    if arr.ndim == 1:
        if arr.size == 16:
            return [arr.reshape(4, 4)]
        if arr.size == 12:
            T = np.eye(4, dtype=np.float32)
            T[:3, :] = arr.reshape(3, 4)
            return [T]
        raise ValueError(f"Unexpected traj format (1D) in {traj_path}: size={arr.size}")
    if arr.ndim == 2:
        if arr.shape[1] == 16:
            return [arr[i].reshape(4, 4) for i in range(arr.shape[0])]
        if arr.shape[1] == 12:
            out = []
            for i in range(arr.shape[0]):
                T = np.eye(4, dtype=np.float32)
                T[:3, :] = arr[i].reshape(3, 4)
                out.append(T)
            return out
    raise ValueError(f"Unexpected traj format in {traj_path}: shape={arr.shape}")


def _backproject_depth_to_world(depth_m, sem_id, Tcw, K4, stride=4, max_depth=10.0):
    """Backproject depth + semantic_id image to world points with labels."""
    K = K4[:3, :3]
    fx, fy, cx, cy = K[0, 0], K[1, 1], K[0, 2], K[1, 2]
    H, W = depth_m.shape

    ys = np.arange(0, H, stride)
    xs = np.arange(0, W, stride)
    grid_y, grid_x = np.meshgrid(ys, xs, indexing="ij")
    u = grid_x.reshape(-1)
    v = grid_y.reshape(-1)

    d = depth_m[v, u]
    lab = sem_id[v, u].astype(np.int32)

    valid = (d > 1e-6) & (d < max_depth)
    u = u[valid]
    v = v[valid]
    d = d[valid]
    lab = lab[valid]

    # camera coordinates
    x = (u - cx) * d / fx
    y = (v - cy) * d / fy
    z = d
    pts_cam = np.stack([x, y, z], axis=1)

    # world coordinates: Xw = R * Xc + t  (Tcw)
    R = Tcw[:3, :3]
    t = Tcw[:3, 3]
    pts_w = (R @ pts_cam.T).T + t[None, :]
    return pts_w, lab


def _voxelize_points_majority(points, labels, voxUnit=0.08):
    """Voxelize points to a grid of size voxUnit, assign semantic label by majority vote.
    Returns (N,7) [x,y,z,r,g,b,label] with rgb filled zeros.
    """
    # quantize to voxel centers
    q = np.floor(points / voxUnit + 0.5).astype(np.int64)  # integer voxel index

    # group by voxel index
    keys = q[:, 0].astype(np.int64), q[:, 1].astype(np.int64), q[:, 2].astype(np.int64)
    key = np.stack(keys, axis=1)

    # sort by key
    order = np.lexsort((key[:, 2], key[:, 1], key[:, 0]))
    key = key[order]
    labels = labels[order]

    out_xyz = []
    out_lab = []

    start = 0
    n = key.shape[0]
    while start < n:
        end = start + 1
        while end < n and np.all(key[end] == key[start]):
            end += 1
        labs = labels[start:end]

        labs_valid = labs[labs > 0]
        if len(labs_valid) == 0:
            maj = 0
        else:
            maj = np.bincount(labs_valid).argmax()

        voxel_center = key[start].astype(np.float32) * voxUnit
        out_xyz.append(voxel_center)
        out_lab.append(maj)
        start = end

    out_xyz = np.stack(out_xyz, axis=0)
    out_lab = np.asarray(out_lab, dtype=np.int32).reshape(-1, 1)

    rgb = np.zeros((out_xyz.shape[0], 3), dtype=np.uint8)
    vox = np.hstack([out_xyz.astype(np.float32), rgb.astype(np.float32), out_lab.astype(np.float32)])  # (N,7)
    return vox


# -----------------------------
# Step 0: build "preprocessed/<scene>.npy"
# -----------------------------

def preprocess_replica_scene(scene_name, replica_root, out_preprocessed,
                             max_frames=-1, stride=4, depth_scale=1000.0, max_depth=10.0,
                             verbose=False):
    """
    Build OccScanNet-like preprocessed voxel list for a Replica scene:
      output: out_preprocessed/<scene>.npy  (N,7): [x,y,z,r,g,b,label]
    """
    scene_dir = os.path.join(replica_root, scene_name)
    if not os.path.isdir(scene_dir):
        return

    traj_path = os.path.join(scene_dir, "traj.txt")
    if not os.path.exists(traj_path):
        raise FileNotFoundError(f"traj.txt not found in {scene_dir}")

    # inputs
    depth_paths = sorted(glob.glob(os.path.join(scene_dir, "depths", "depth*.png")))
    sem_paths = sorted(glob.glob(os.path.join(scene_dir, "semantic_ids", "semantic_id*.png")))

    if len(depth_paths) == 0 or len(sem_paths) == 0:
        raise FileNotFoundError(f"depths/ or semantic_ids/ missing or empty in {scene_dir}")

    # align by index (assume same count/order)
    n = min(len(depth_paths), len(sem_paths))
    depth_paths = depth_paths[:n]
    sem_paths = sem_paths[:n]

    Tcw_list = _load_traj_as_Tcw_list(traj_path)
    n = min(n, len(Tcw_list))
    depth_paths = depth_paths[:n]
    sem_paths = sem_paths[:n]
    Tcw_list = Tcw_list[:n]

    if max_frames is not None and max_frames > 0:
        n = min(n, max_frames)
        depth_paths = depth_paths[:n]
        sem_paths = sem_paths[:n]
        Tcw_list = Tcw_list[:n]

    K4, file_depth_scale = _try_load_intrinsics(scene_dir)

    all_pts = []
    all_lab = []

    it = range(n)
    if verbose:
        it = tqdm(it, desc=f"[preprocess] {scene_name}", leave=False)

    for i in it:
        # depth
        d16 = Image.open(depth_paths[i]).convert("I;16")
        d16 = np.array(d16).astype(np.float32)

        use_scale = float(depth_scale)
        if (use_scale is None) or (use_scale <= 0):
            if file_depth_scale is None:
                raise ValueError("depth_scale not provided and cam_params.json has no 'scale'")
            use_scale = float(file_depth_scale)

        depth_m = d16 / use_scale

        # semantic id (assume 8-bit or 16-bit; keep as int)
        sem = Image.open(sem_paths[i])
        sem = np.array(sem).astype(np.int32)

        pts_w, lab = _backproject_depth_to_world(
            depth_m=depth_m,
            sem_id=sem,
            Tcw=Tcw_list[i],
            K4=K4,
            stride=stride,
            max_depth=max_depth,
        )
        if pts_w.shape[0] == 0:
            continue
        all_pts.append(pts_w)
        all_lab.append(lab)

    if len(all_pts) == 0:
        raise RuntimeError(f"No valid backprojected points for scene {scene_name}")

    all_pts = np.concatenate(all_pts, axis=0)
    all_lab = np.concatenate(all_lab, axis=0)

    # voxelize -> (N,7)
    vox = _voxelize_points_majority(all_pts, all_lab, voxUnit=0.08)

    os.makedirs(out_preprocessed, exist_ok=True)
    np.save(os.path.join(out_preprocessed, f"{scene_name}.npy"), vox)

    return True


# -----------------------------
# multiprocessing worker
# -----------------------------

def worker_preprocess_scene(scene_name):
    return preprocess_replica_scene(
        scene_name=scene_name,
        replica_root=_G_REPLICA_ROOT,
        out_preprocessed=_G_OUT_PREPROCESSED,
        max_frames=_G_MAX_FRAMES,
        stride=_G_STRIDE,
        depth_scale=_G_DEPTH_SCALE,
        max_depth=_G_MAX_DEPTH,
        verbose=False,
    )


# -----------------------------
# Main
# -----------------------------

def parse_args():
    p = argparse.ArgumentParser("Prepare Replica -> preprocessed voxels (Step 0 only)")
    p.add_argument("--replica_root", type=str, required=True,
                   help="Path to Replica-SLAM root, containing office0/room0/..")
    p.add_argument("--out_root", type=str, required=True,
                   help="Output root. Will create preprocessed/")
    p.add_argument("--nproc", type=int, default=1)
    p.add_argument("--max_frames", type=int, default=-1,
                   help="Use first N frames per scene for building global preprocessed voxels; -1 means all.")
    p.add_argument("--stride", type=int, default=4,
                   help="Pixel stride for backprojection when building global preprocessed voxels.")
    p.add_argument("--depth_scale", type=float, default=-1,
                   help="depth_m = depth_png / depth_scale. Common: 1000 if png in mm.")
    p.add_argument("--max_depth", type=float, default=10.0)
    return p.parse_args()


def main():
    args = parse_args()
    replica_root = args.replica_root
    out_root = args.out_root

    out_preprocessed = os.path.join(out_root, "preprocessed")
    os.makedirs(out_preprocessed, exist_ok=True)

    scene_list = sorted([d for d in os.listdir(replica_root)
                         if os.path.isdir(os.path.join(replica_root, d))])

    global _G_REPLICA_ROOT, _G_OUT_PREPROCESSED, _G_MAX_FRAMES, _G_STRIDE, _G_DEPTH_SCALE, _G_MAX_DEPTH
    _G_REPLICA_ROOT = replica_root
    _G_OUT_PREPROCESSED = out_preprocessed
    _G_MAX_FRAMES = args.max_frames
    _G_STRIDE = args.stride
    _G_DEPTH_SCALE = args.depth_scale
    _G_MAX_DEPTH = args.max_depth

    # Run the selected preprocessing tasks.
    track_parallel_progress(worker_preprocess_scene, scene_list, nproc=args.nproc)
    print("===== Finish Step 0 (preprocessed) =====")


if __name__ == "__main__":
    main()