Room-scale/calibration/scripts/visualize/visualize_case.py

#!/usr/bin/env python3
"""Visualize a synchronized ZED + Homie + OptiTrack/SMPL-X data case in Rerun."""

from __future__ import annotations

import argparse
import csv
import json
import os
import socket
import struct
import subprocess
import time
import xml.etree.ElementTree as ET
from pathlib import Path

import cv2
import numpy as np
import rerun as rr
import rerun.blueprint as rrb
import yaml

HERE = Path(__file__).resolve().parent
DEFAULT_CASE = Path("/mnt/Datasets/2026-06-27-12:26:35")
HOMIE_NAMES = ("front_left", "front_right", "back_left", "back_right")
HOMIE_VIDEO = {
    "front_left": "homie_front_left.mp4",
    "front_right": "homie_front_right.mp4",
    "back_left": "homie_back_left.mp4",
    "back_right": "homie_back_right.mp4",
}
HOMIE_CAMCHAIN = {
    "front_left": "cam0", "front_right": "cam1",
    "back_left": "cam2", "back_right": "cam3",
}


def parse_args() -> argparse.Namespace:
    p = argparse.ArgumentParser(formatter_class=argparse.ArgumentDefaultsHelpFormatter)
    p.add_argument("case", nargs="?", type=Path, default=DEFAULT_CASE)
    p.add_argument("--zed-calib", type=Path, default=HERE / "config/zed_poses_0622.json")
    p.add_argument("--homie-calib", type=Path, default=HERE / "config/homie_rig_calib.json")
    p.add_argument("--homie-camchain", type=Path, default=HERE / "config/homie-camchain.yaml")
    p.add_argument("--zed", default="0,1,2,3,4,5,6,7", help="ZED camera indices, or empty")
    p.add_argument("--homie", default="front_left", help="Homie camera names, or empty")
    p.add_argument("--optitrack-calib", type=Path, default=HERE / "config/Calibration-Exported.mcal")
    p.add_argument("--assets-dir", type=Path,
                   default=Path("/home/acerobotics/Datasets/mesh_alignmend/aligned_textured"),
                   help="room OBJ asset directory (flat *.obj or <name>/mesh.obj layout)")
    p.add_argument("--no-room-assets", dest="no_room_assets", action="store_true", default=False,
                   help="do not load room assets")
    p.add_argument("--room-assets", dest="no_room_assets", action="store_false",
                   help="explicitly enable room asset loading")
    p.add_argument("--enable-room-alignment", dest="enable_room_alignment", action="store_true", default=True,
                   help="enable floor/center alignment and CLI room offsets (default)")
    p.add_argument("--no-room-alignment", dest="enable_room_alignment", action="store_false",
                   help="disable floor/center alignment and CLI room offsets")
    p.add_argument("--enable-room-controls", dest="enable_room_controls", action="store_true", default=True,
                   help="launch live room translation/rotation controls (default)")
    p.add_argument("--no-room-controls", dest="enable_room_controls", action="store_false",
                   help="disable live room transform controls")
    p.add_argument("--asset-floor-z", type=float, default=-1.62, help="floor Z in the source Z-up OBJ coordinates")
    p.add_argument("--asset-translation", type=float, nargs=3, default=(0.12, -0.52, 0.79), metavar=("DX", "DY", "DZ"),
                   help="additional room translation in OptiTrack world meters")
    p.add_argument("--asset-rotation-deg", type=float, nargs=3, default=(0.0, 62.0, 0.0), metavar=("RX", "RY", "RZ"),
                   help="additional room rotation about world X/Y/Z in degrees")
    p.add_argument("--no-auto-asset-align", action="store_true", help="disable horizontal room/camera center alignment")
    p.add_argument("--start", type=float, default=0.0, help="start time in OptiTrack seconds")
    p.add_argument("--duration", type=float, default=None)
    p.add_argument("--mesh-fps", type=float, default=15.0,
                   help="(unused; SMPL is logged as a single static pose at --start)")
    p.add_argument("--no-video", action="store_true")
    p.add_argument("--full-video", action="store_true", help="embed complete MP4 assets instead of one RGB frame per camera")
    p.add_argument("--no-smpl", action="store_true")
    p.add_argument("--save", type=Path, help="write an .rrd instead of spawning the viewer")
    p.add_argument("--connect", metavar="URL", help="connect to an existing Rerun server, e.g. rerun+http://127.0.0.1:9876/proxy")
    p.add_argument("--grpc-port", type=int, default=9876, help="Rerun data proxy port (only with --connect)")
    p.add_argument("--web-port", type=int, default=9090, help="headless Web Viewer HTTP port")
    p.add_argument("--control-port", type=int, default=None, help="room-alignment UI port; default web-port + 2")
    p.add_argument("--server-memory", default="8GiB", help="Rerun proxy cache limit")
    return p.parse_args()


def kill_orphaned_viewers() -> None:
    """Terminate leftover rerun servers from previous runs of this script.

    They are detached (start_new_session), so they outlive a crashed or
    Ctrl+C'd parent and then squat on the ports. A process serving our own
    /tmp/multicam_optitrack_smpl_*.rrd is killed unless some ancestor is a
    still-running visualize_case.py (i.e. an active sibling session).
    """
    import signal

    def ppid_of(pid: int) -> int:
        try:
            return int(Path(f"/proc/{pid}/stat").read_text().rsplit(")", 1)[1].split()[1])
        except (OSError, IndexError, ValueError):
            return 0

    def has_live_owner(pid: int) -> bool:
        seen = 0
        while pid > 1 and seen < 20:
            try:
                argv = Path(f"/proc/{pid}/cmdline").read_bytes().split(b"\0")
            except OSError:
                return False
            if pid != os.getpid() and any(a.endswith(b"visualize_case.py") for a in argv):
                return True
            pid, seen = ppid_of(pid), seen + 1
        return False

    for pid_dir in Path("/proc").glob("[0-9]*"):
        pid = int(pid_dir.name)
        if pid == os.getpid():
            continue
        try:
            argv = (pid_dir / "cmdline").read_bytes().split(b"\0")
        except OSError:
            continue
        # exact-argument match so shells whose command text merely quotes
        # these strings are never touched
        serves_our_rrd = any(a.startswith(b"/tmp/multicam_optitrack_smpl_")
                             and a.endswith(b".rrd") for a in argv)
        if (serves_our_rrd and b"--serve-web" in argv
                and not has_live_owner(pid)):
            print(f"Stopping orphaned Rerun viewer (pid {pid})")
            try:
                os.kill(pid, signal.SIGTERM)
            except OSError:
                pass
    time.sleep(0.5)


def start_headless_server(recording: Path, web_port: int, server_memory: str) -> subprocess.Popen:
    """Start the web viewer on a completed RRD; no transient gRPC stream involved."""
    executable = Path(__file__).resolve().parent / ".venv/bin/rerun"
    if not executable.is_file():
        raise FileNotFoundError(f"Rerun CLI not found at {executable}; install requirements in .venv")
    with socket.socket() as check:
        if check.connect_ex(("127.0.0.1", web_port)) == 0:
            kill_orphaned_viewers()
    with socket.socket() as check:
        if check.connect_ex(("127.0.0.1", web_port)) == 0:
            raise RuntimeError(f"Port {web_port} is already in use. Re-run with --web-port {web_port + 1}")
    # Rerun's web app still fetches data through this proxy. Keep it predictable for SSH.
    proxy_port = web_port + 1
    with socket.socket() as check:
        if check.connect_ex(("127.0.0.1", proxy_port)) == 0:
            raise RuntimeError(f"Proxy port {proxy_port} is already in use. Choose another --web-port")
    proc = subprocess.Popen(
        [str(executable), str(recording), "--serve-web", "--bind", "127.0.0.1",
         "--port", str(proxy_port), "--web-viewer-port", str(web_port),
         "--server-memory-limit", server_memory],
        start_new_session=True,
    )
    deadline = time.monotonic() + 15.0
    while time.monotonic() < deadline:
        if proc.poll() is not None:
            raise RuntimeError("Rerun Web Viewer exited during startup; run .venv/bin/rerun --serve-web to inspect its error")
        try:
            with socket.create_connection(("127.0.0.1", web_port), timeout=0.2):
                return proc
        except OSError:
            time.sleep(0.1)
    proc.terminate()
    raise TimeoutError(f"Rerun Web Viewer did not open port {web_port}")


def rotation_xyz_deg(rx: float, ry: float, rz: float) -> np.ndarray:
    x, y, z = np.radians([rx, ry, rz])
    cx, sx, cy, sy, cz, sz = np.cos(x), np.sin(x), np.cos(y), np.sin(y), np.cos(z), np.sin(z)
    Rx = np.array([[1, 0, 0], [0, cx, -sx], [0, sx, cx]])
    Ry = np.array([[cy, 0, sy], [0, 1, 0], [-sy, 0, cy]])
    Rz = np.array([[cz, -sz, 0], [sz, cz, 0], [0, 0, 1]])
    return Rz @ Ry @ Rx


def launch_alignment_ui(base_translation: np.ndarray, room_R: np.ndarray, proxy_port: int,
                        control_port: int, recording_id: str):
    import gradio as gr
    stream = rr.RecordingStream("multicam_optitrack_smpl", recording_id=recording_id)
    stream.connect_grpc(f"rerun+http://127.0.0.1:{proxy_port}/proxy")

    counter = [0]

    def update(dx: float, dy: float, dz: float, rx: float, ry: float, rz: float) -> str:
        delta = np.array([dx, dy, dz], float)
        total = base_translation + delta
        total_R = rotation_xyz_deg(rx, ry, rz) @ room_R
        rr.log("world/room", rr.Transform3D(translation=total, mat3x3=total_R), static=True, recording=stream)
        stream.flush()
        counter[0] += 1
        return (f"已发送更新 #{counter[0]}\n\n"
                f"总平移: [{total[0]:.3f}, {total[1]:.3f}, {total[2]:.3f}] m  "
                f"附加旋转: [{rx:.1f}, {ry:.1f}, {rz:.1f}]°\n\n"
                f"启动参数: `--asset-translation {dx:.3f} {dy:.3f} {dz:.3f} "
                f"--asset-rotation-deg {rx:.1f} {ry:.1f} {rz:.1f}`")

    with gr.Blocks(title="Room alignment") as demo:
        gr.Markdown("# 房间 Mesh 平移微调\n滑块表示叠加在自动对齐结果上的偏移（米）。松开滑块后更新 Rerun。")
        x = gr.Slider(-5, 5, value=0, step=0.01, label="X 平移 / m")
        y = gr.Slider(-5, 5, value=0, step=0.01, label="Y 平移 / m")
        z = gr.Slider(-5, 5, value=0, step=0.01, label="Z 平移 / m")
        with gr.Row():
            rx = gr.Slider(-180, 180, value=0, step=1, label="绕 X 旋转 / °")
            ry = gr.Slider(-180, 180, value=0, step=1, label="绕 Y 旋转 / °")
            rz = gr.Slider(-180, 180, value=0, step=1, label="绕 Z 旋转 / °")
        controls = [x, y, z, rx, ry, rz]
        status = gr.Markdown(update(0, 0, 0, 0, 0, 0))
        for slider in controls:
            slider.release(update, inputs=controls, outputs=status, show_progress="hidden")
        gr.Button("应用数值", variant="primary").click(update, inputs=controls, outputs=status, show_progress="hidden")
    demo.launch(server_name="127.0.0.1", server_port=control_port, prevent_thread_lock=True, quiet=True)
    return demo, stream


def load_obj(path: Path) -> tuple[np.ndarray, np.ndarray]:
    vertices, faces = [], []
    with path.open() as f:
        for line in f:
            if line.startswith("v "):
                vertices.append([float(x) for x in line.split()[1:4]])
            elif line.startswith("f "):
                idx = [int(x.split("/")[0]) - 1 for x in line.split()[1:]]
                for i in range(1, len(idx) - 1):
                    faces.append([idx[0], idx[i], idx[i + 1]])
    return np.asarray(vertices, np.float32), np.asarray(faces, np.uint32)


def obj_collection_bounds(paths: list[Path]) -> tuple[np.ndarray, np.ndarray]:
    lo, hi = np.full(3, np.inf), np.full(3, -np.inf)
    for path in paths:
        with path.open(errors="ignore") as f:
            for line in f:
                if line.startswith("v "):
                    v = np.fromstring(line[2:], sep=" ", count=3)
                    if len(v) == 3:
                        lo, hi = np.minimum(lo, v), np.maximum(hi, v)
    if not np.isfinite(lo).all():
        raise ValueError("No OBJ vertices found for room alignment")
    return lo, hi


def open_pc2(path: Path) -> tuple[np.memmap, float]:
    with path.open("rb") as f:
        if f.read(12) != b"POINTCACHE2\0":
            raise ValueError(f"Not a POINTCACHE2 file: {path}")
        version, points, start, rate, frames = struct.unpack("<iiffi", f.read(20))
    if version != 1 or path.stat().st_size != 32 + frames * points * 12:
        raise ValueError(f"Invalid PC2 header/size: {path}")
    return np.memmap(path, dtype="<f4", mode="r", offset=32, shape=(frames, points, 3)), rate


def read_homie_poses(path: Path) -> tuple[np.ndarray, np.ndarray, np.ndarray]:
    data = np.genfromtxt(path, delimiter=",", names=True, skip_header=1)
    frames = data["frame"].astype(np.int64)
    pos_m = np.column_stack([data["Homie_px"], data["Homie_py"], data["Homie_pz"]]) / 1000.0
    quat = np.column_stack([data["Homie_qx"], data["Homie_qy"], data["Homie_qz"], data["Homie_qw"]])
    return frames, pos_m, quat


def quat_xyzw_to_mat(q: np.ndarray) -> np.ndarray:
    q = q / np.linalg.norm(q)
    x, y, z, w = q
    return np.array([
        [1-2*(y*y+z*z), 2*(x*y-z*w), 2*(x*z+y*w)],
        [2*(x*y+z*w), 1-2*(x*x+z*z), 2*(y*z-x*w)],
        [2*(x*z-y*w), 2*(y*z+x*w), 1-2*(x*x+y*y)],
    ])


def euler_rotation(pitch: float, yaw: float, roll: float) -> np.ndarray:
    p, y, r = np.radians([pitch, yaw, roll])
    cp, sp, cy, sy, cr, sr = np.cos(p), np.sin(p), np.cos(y), np.sin(y), np.cos(r), np.sin(r)
    return np.array([[cr*cy, cr*sy*sp-sr*cp, cr*sy*cp+sr*sp],
                     [sr*cy, sr*sy*sp+cr*cp, sr*sy*cp-cr*sp],
                     [-sy, cy*sp, cy*cp]])


def homie_rectification(camchain_path: Path) -> tuple[dict[str, np.ndarray], np.ndarray]:
    cc = yaml.safe_load(camchain_path.read_text())
    # Matches the parameters recorded in this case's rectify_report.json.
    out, fov, zoom = 1024, 90.0, 0.78125
    focal = out / (2 * np.tan(np.radians(fov) / 2)) * zoom
    K = np.array([[focal, 0, out/2], [0, focal, out/2], [0, 0, 1.0]])
    view = euler_rotation(-5.0, 0.0, 0.0)
    result = {}
    pairs = (("cam0", "front_left", "cam1", "front_right"),
             ("cam2", "back_left", "cam3", "back_right"))
    for lk, ln, rk, rn in pairs:
        def kd(key: str):
            c = cc[key]; fx, fy, cx, cy = c["intrinsics"]
            return np.array([[fx, 0, cx], [0, fy, cy], [0, 0, 1.0]]), np.asarray(c["distortion_coeffs"])
        K0, D0 = kd(lk); K1, D1 = kd(rk)
        T = np.asarray(cc[rk]["T_cn_cnm1"], float)
        size = tuple(map(int, cc[lk]["resolution"]))
        R1, R2, *_ = cv2.fisheye.stereoRectify(K0, D0, K1, D1, size, T[:3, :3], T[:3, 3],
                                                flags=cv2.CALIB_ZERO_DISPARITY, newImageSize=size,
                                                balance=0.0, fov_scale=1.0)
        result[ln], result[rn] = view @ R1, view @ R2
    return result, K


def log_video(path: str, video_path: Path, origin_s: float) -> None:
    asset = rr.AssetVideo(path=video_path)
    rr.log(path, asset, static=True)
    stamps = asset.read_frame_timestamps_nanos()
    rr.send_columns(path,
        indexes=[rr.TimeColumn("time", duration=origin_s + stamps * 1e-9)],
        columns=rr.VideoFrameReference.columns_nanos(stamps))


def log_video_frame(path: str, video_path: Path, origin_s: float, requested_time_s: float,
                    flip_vertical: bool = False) -> None:
    """Log one JPEG frame, avoiding transfer of the complete MP4 asset."""
    sample_time = max(requested_time_s, origin_s)
    cap = cv2.VideoCapture(str(video_path))
    cap.set(cv2.CAP_PROP_POS_MSEC, max(0.0, sample_time - origin_s) * 1000.0)
    ok, frame = cap.read()
    cap.release()
    if not ok:
        raise RuntimeError(f"Could not decode RGB frame from {video_path} at {sample_time - origin_s:.3f}s")
    if flip_vertical:
        frame = cv2.flip(frame, 0)
    ok, encoded = cv2.imencode(".jpg", frame, [cv2.IMWRITE_JPEG_QUALITY, 90])
    if not ok:
        raise RuntimeError(f"Could not JPEG-encode frame from {video_path}")
    rr.set_time("time", duration=sample_time)
    rr.log(path, rr.EncodedImage(contents=encoded.tobytes(), media_type="image/jpeg"))


def rotation_midpoint(a: np.ndarray, b: np.ndarray) -> np.ndarray:
    """Closest proper rotation to the equal-weight mean of two rotations."""
    u, _, vt = np.linalg.svd(a + b)
    d = np.eye(3); d[-1, -1] = np.linalg.det(u @ vt)
    return u @ d @ vt


def load_optitrack_cameras(path: Path) -> dict[int, dict]:
    root = ET.fromstring(path.read_bytes().decode("utf-16"))
    gl_to_cv = np.diag([1.0, -1.0, -1.0])
    cameras = {}
    for cam in root.findall(".//Camera"):
        props, intr, extr = cam.find("Properties"), cam.find("IntrinsicStandardCameraModel"), cam.find("Extrinsic")
        if props is None or intr is None or extr is None:
            continue
        cid = int(props.attrib["CameraID"])
        fx, fy = float(intr.attrib["HorizontalFocalLength"]), float(intr.attrib["VerticalFocalLength"])
        cx, cy = float(intr.attrib["LensCenterX"]), float(intr.attrib["LensCenterY"])
        R = np.array([float(extr.attrib[f"OrientMatrix{i}"]) for i in range(9)]).reshape(3, 3) @ gl_to_cv
        t = np.array([float(extr.attrib[k]) for k in ("X", "Y", "Z")]) * 1000.0
        T = np.eye(4); T[:3, :3], T[:3, 3] = R, t
        cameras[cid] = {"T_c2w": T, "K": np.array([[fx, 0, cx], [0, fy, cy], [0, 0, 1.0]]),
                        "width": 2048, "height": 1088, "serial": cam.attrib.get("Serial", "")}
    return cameras


def log_camera_static(path: str, T_c2w: np.ndarray, K: np.ndarray, width: int, height: int,
                      color: tuple[int, int, int]) -> None:
    rr.log(path, rr.Transform3D(translation=T_c2w[:3, 3] / 1000.0, mat3x3=T_c2w[:3, :3]), static=True)
    rr.log(path, rr.Pinhole(image_from_camera=K, width=width, height=height,
                            camera_xyz=rr.ViewCoordinates.RDF, image_plane_distance=0.32,
                            color=color), static=True)


def blueprint(zed: list[str], homie: list[str]) -> rrb.Blueprint:
    video_views = [rrb.Spatial2DView(name=f"ZED {c}", origin=f"world/cameras/zed/cam{c}") for c in zed]
    video_views += [rrb.Spatial2DView(name=f"Homie {c}", origin=f"world/cameras/homie/{c}") for c in homie]
    videos = rrb.Grid(*video_views, grid_columns=2) if video_views else rrb.TextLogView()
    return rrb.Blueprint(rrb.Horizontal(
        rrb.Spatial3DView(name="All cameras + SMPL", origin="world", contents=["world/**"],
                          background=[41, 43, 38]),  # matches calibration_sz.png backdrop
        videos, column_shares=[2, 1]), collapse_panels=True)


def main() -> None:
    args = parse_args(); case = args.case.resolve()
    server_proc = None
    alignment_ui = alignment_stream = None
    room_base_translation = room_rotation = None
    recording_id = "multicam-optitrack-smpl-room-control"
    web_rrd = None
    zed_ids = [x.strip() for x in args.zed.split(",") if x.strip()]
    homie_ids = [x.strip() for x in args.homie.split(",") if x.strip()]
    unknown = set(homie_ids) - set(HOMIE_NAMES)
    if unknown: raise ValueError(f"Unknown Homie cameras: {sorted(unknown)}")
    if args.save and args.connect:
        raise ValueError("--save and --connect are mutually exclusive")
    rr.init("multicam_optitrack_smpl", recording_id=recording_id, spawn=False)
    if args.save:
        rr.save(args.save)
    elif args.connect:
        rr.connect_grpc(args.connect)
    else:
        web_rrd = Path(f"/tmp/multicam_optitrack_smpl_{os.getpid()}.rrd")
        rr.save(web_rrd)
    rr.log("world", rr.ViewCoordinates.RIGHT_HAND_Y_UP, static=True)
    rr.log("world/origin", rr.TransformAxes3D(axis_length=0.5), static=True)

    zc = json.loads(args.zed_calib.read_text())["cameras"]

    if not args.no_room_assets and args.assets_dir.is_dir():
        # Source meshes are globally placed in a Z-up reconstruction. Convert
        # (x,y,z)_src -> (x,z,-y)_world and lift their source floor to Y=0.
        axis_R = np.array([[1, 0, 0], [0, 0, 1], [0, -1, 0]], dtype=float)
        room_R = axis_R.copy()
        excluded_assets = {"scene.obj", "table_cook.obj"}
        # New layout: aligned_textured/<name>/mesh.glb (textures embedded;
        # Asset3D cannot resolve an OBJ's external .mtl/.png). Fall back to
        # <name>/mesh.obj, then to the legacy flat *.obj layout.
        parts = sorted(args.assets_dir.glob("*/mesh.glb"))
        if not parts:
            parts = sorted(args.assets_dir.glob("*/mesh.obj"))
        if not parts:
            parts = sorted(p for p in args.assets_dir.glob("*.obj") if p.name not in excluded_assets)
        objects = parts
        translation = np.zeros(3)
        if args.enable_room_alignment:
            room_R = rotation_xyz_deg(*args.asset_rotation_deg) @ axis_R
            translation = np.array([0.0, -args.asset_floor_z, 0.0]) + np.asarray(args.asset_translation)
        if args.enable_room_alignment and objects and not args.no_auto_asset_align:
            # bounds parser reads OBJ text; GLB entries use their sibling OBJ
            bounds_paths = [p.with_suffix(".obj") if p.suffix == ".glb" else p for p in objects]
            lo, hi = obj_collection_bounds(bounds_paths)
            room_center_world = room_R @ ((lo + hi) * 0.5)
            camera_positions = np.array([np.asarray(c["T_c2w"], float)[:3, 3] / 1000.0 for c in zc.values()])
            camera_center = (camera_positions.min(axis=0) + camera_positions.max(axis=0)) * 0.5
            translation[[0, 2]] += camera_center[[0, 2]] - room_center_world[[0, 2]]
            print(f"Auto room alignment translation: {translation.round(3).tolist()} m")
        rr.log("world/room", rr.Transform3D(translation=translation, mat3x3=room_R), static=True)
        if args.enable_room_controls:
            room_base_translation, room_rotation = translation.copy(), room_R.copy()
        palette = [(230, 92, 92), (239, 158, 64), (230, 205, 80), (86, 190, 116),
                   (65, 185, 205), (78, 135, 225), (145, 105, 210), (215, 95, 175),
                   (155, 115, 82), (120, 175, 155)]
        for i, obj in enumerate(objects):
            # <name>/mesh.glb|obj carries its own texture: log untinted.
            # Legacy untextured flat OBJs keep the palette tint.
            textured = obj.name in ("mesh.glb", "mesh.obj")
            name = obj.parent.name if textured else obj.stem
            if textured:
                rr.log(f"world/room/objects/{name}", rr.Asset3D(path=obj), static=True)
            else:
                rr.log(f"world/room/objects/{name}",
                       rr.Asset3D(path=obj, albedo_factor=palette[i % len(palette)]), static=True)

    zed_color = (60, 220, 90)
    for cid in zed_ids:
        c = zc[cid]; path = f"world/cameras/zed/cam{cid}"
        log_camera_static(path, np.asarray(c["T_c2w"]), np.asarray(c["K"]), c["width"], c["height"], zed_color)

    for cid, c in load_optitrack_cameras(args.optitrack_calib).items():
        log_camera_static(f"world/cameras/optitrack/cam{cid}", c["T_c2w"], c["K"], c["width"], c["height"], (65, 135, 255))

    rig = json.loads(args.homie_calib.read_text())["T_homie_cam"]
    Rrect, Kh = homie_rectification(args.homie_camchain)
    frames, positions, quats = read_homie_poses(case / "objects/rb_Homie.csv")
    end = np.inf if args.duration is None else args.start + args.duration

    def homie_cam_T(Twh: np.ndarray, cam: str) -> np.ndarray:
        if cam == "front_left":
            Tl = Twh @ np.asarray(rig["front_left"], float)
            Tr = Twh @ np.asarray(rig["front_right"], float)
            Tl[:3, :3] = Tl[:3, :3] @ Rrect["front_left"].T
            Tr[:3, :3] = Tr[:3, :3] @ Rrect["front_right"].T
            T = np.eye(4)
            T[:3, 3] = (Tl[:3, 3] + Tr[:3, 3]) * 0.5
            T[:3, :3] = rotation_midpoint(Tl[:3, :3], Tr[:3, :3])
            return T
        T = Twh @ np.asarray(rig[cam], float)
        T[:3, :3] = T[:3, :3] @ Rrect[cam].T
        return T

    if args.full_video:
        # animated rig: poses follow the timeline alongside the videos
        keep = (frames / 60.0 >= args.start) & (frames / 60.0 <= end)
        for frame, pos, quat in zip(frames[keep], positions[keep], quats[keep]):
            rr.set_time("time", duration=float(frame) / 60.0)
            Rwh = quat_xyzw_to_mat(quat)
            rr.log("world/optitrack/homie_rigid_body", rr.Transform3D(translation=pos, mat3x3=Rwh))
            rr.log("world/optitrack/homie_rigid_body", rr.TransformAxes3D(axis_length=0.25))
            Twh = np.eye(4); Twh[:3, :3], Twh[:3, 3] = Rwh, pos * 1000.0
            for cam in homie_ids:
                T = homie_cam_T(Twh, cam)
                rr.log(f"world/cameras/homie/{cam}", rr.Transform3D(translation=T[:3, 3]/1000.0, mat3x3=T[:3, :3]))
    else:
        # static snapshot at --start, consistent with the static human and the
        # single logged video frames
        i = int(np.argmin(np.abs(frames / 60.0 - args.start)))
        Rwh = quat_xyzw_to_mat(quats[i])
        rr.log("world/optitrack/homie_rigid_body",
               rr.Transform3D(translation=positions[i], mat3x3=Rwh), static=True)
        rr.log("world/optitrack/homie_rigid_body", rr.TransformAxes3D(axis_length=0.25), static=True)
        Twh = np.eye(4); Twh[:3, :3], Twh[:3, 3] = Rwh, positions[i] * 1000.0
        for cam in homie_ids:
            T = homie_cam_T(Twh, cam)
            rr.log(f"world/cameras/homie/{cam}",
                   rr.Transform3D(translation=T[:3, 3]/1000.0, mat3x3=T[:3, :3]), static=True)
    for cam in homie_ids:
        rr.log(f"world/cameras/homie/{cam}", rr.Pinhole(image_from_camera=Kh, width=1024, height=1024,
            camera_xyz=rr.ViewCoordinates.RDF, image_plane_distance=0.24, color=(255, 105, 55)), static=True)

    if not args.no_video:
        export = json.loads((case / "videos/export_info.json").read_text())
        origins = {v["file"]: float(v["start_optitrack_time_s"]) for v in export["videos"]}
        for cid in zed_ids:
            name = f"zed_cam{cid}.mp4"; path = f"world/cameras/zed/cam{cid}/rgb"
            if args.full_video: log_video(path, case/"videos"/name, origins[name])
            else: log_video_frame(path, case/"videos"/name, origins[name], args.start, flip_vertical=True)
        for cam in homie_ids:
            name = HOMIE_VIDEO[cam]; path = f"world/cameras/homie/{cam}/rgb"
            if args.full_video: log_video(path, case/"videos"/name, origins[name])
            else: log_video_frame(path, case/"videos"/name, origins[name], args.start)

    if not args.no_smpl:
        # Static single pose at --start: the Homie rig does not follow the
        # human, so animating the SMPL mesh only inflates the recording.
        _, faces = load_obj(case / "viz/human/poses_template.obj")
        verts, source_fps = open_pc2(case / "viz/human/poses_motion.pc2")
        i = min(len(verts) - 1, max(0, int(round(args.start * source_fps))))
        rr.log("world/human/smplx", rr.Mesh3D(vertex_positions=verts[i], triangle_indices=faces,
                                               vertex_colors=(115, 170, 255, 210)), static=True)
    rr.send_blueprint(blueprint(zed_ids, homie_ids))
    print("Rerun logging complete.")
    if web_rrd is not None:
        rr.disconnect()
        server_proc = start_headless_server(web_rrd, args.web_port, args.server_memory)
        print(f"Headless Web Viewer loaded {web_rrd}.")
        proxy_port = args.web_port + 1
        control_port = args.control_port
        if control_port is None:  # first free port from web_port+2
            for candidate in range(args.web_port + 2, args.web_port + 22):
                with socket.socket() as check:
                    if check.connect_ex(("127.0.0.1", candidate)) != 0:
                        control_port = candidate
                        break
            else:
                raise RuntimeError("No free control port found; pass --control-port")
        if room_base_translation is not None:
            alignment_ui, alignment_stream = launch_alignment_ui(
                room_base_translation, room_rotation, proxy_port, control_port, recording_id)
        print("On your LOCAL machine run:")
        forwards = f"-L {args.web_port}:127.0.0.1:{args.web_port} -L {proxy_port}:127.0.0.1:{proxy_port}"
        if alignment_ui is not None:
            forwards += f" -L {control_port}:127.0.0.1:{control_port}"
        print(f"  ssh -N {forwards} acerobotics@SERVER")
        print("Then open:")
        print(f"  http://127.0.0.1:{args.web_port}?url=rerun%3A%2F%2Flocalhost%3A{proxy_port}%2Fproxy")
        if alignment_ui is not None:
            print(f"Room alignment controls: http://127.0.0.1:{control_port}")
    if server_proc is not None:
        print("Web Viewer is ready. Keep this process running; press Ctrl+C after viewing.")
        try:
            server_proc.wait()
        except KeyboardInterrupt:
            print("\nStopping Web Viewer...")
            if alignment_ui is not None:
                alignment_ui.close()
            server_proc.terminate()
            try:
                server_proc.wait(timeout=5)
            except subprocess.TimeoutExpired:
                server_proc.kill()


if __name__ == "__main__":
    main()