Blame for reconstruction/scripts/viz_object_motion.py

#!/usr/bin/env python3
"""Render one overlay video: tracked 3D object + HaWoR hands on the RGB clip.

Uses the optimized camera-frame layout when present (hand-anchored scale),
otherwise the combined layout from Fast-SAM3D.

  python viz_object_motion.py --video-dir /path/to/runs/<seq>

Writes: <video-dir>/obj_tracking_out/<object>/object_motion.mp4
"""

from __future__ import annotations

import argparse
import json
import os
import sys

import cv2
import numpy as np
import torch
from scipy.spatial.transform import Rotation as R
from pytorch3d.io import load_objs_as_meshes
from pytorch3d.structures import Meshes
from pytorch3d.renderer import TexturesVertex

from run_project_mesh_combined import (
    DirectoryVideoReader,
    _parse_rgb,
    make_renderer,
    read_video_frame,
    render_combined,
    render_mesh,
    transform_to_camera_frame,
)


def parse_args():
    p = argparse.ArgumentParser(
        description="One MP4 of 3D object + hand motion overlaid on the demo video.",
        formatter_class=argparse.ArgumentDefaultsHelpFormatter,
    )
    p.add_argument("--video-dir", required=True, help="Pipeline run directory")
    p.add_argument("--object", default=None, help="Object id (default: config.json)")
    p.add_argument("--video", default=None, help="RGB source (mp4 or all_frames/). Default: clip.mp4")
    p.add_argument("--mesh", default=None, help="Object .obj (default: SAM-3D mesh on ref frame)")
    p.add_argument("--json", default=None, help="Layout JSON (prefers optimized camera-frame)")
    p.add_argument("--hand-meshes", default=None, help="HaWoR all_hand_meshes.npz")
    p.add_argument(
        "--hands",
        default="auto",
        choices=["auto", "left", "right", "both", "none"],
        help="Hands to overlay. auto = config.json anchor_hand (one side). "
             "both also drops near-camera ghost meshes (unused HaWoR side).",
    )
    p.add_argument("--output", default=None, help="Output mp4 path")
    p.add_argument("--fps", type=float, default=30.0)
    p.add_argument(
        "--object-color",
        default="0.15,0.85,0.45",
        help="Solid overlay RGB in 0-1 as r,g,b",
    )
    p.add_argument("--left-hand-color", default="0.95,0.35,0.35")
    p.add_argument("--right-hand-color", default="0.35,0.45,0.95")
    p.add_argument(
        "--alpha",
        type=float,
        default=0.72,
        help="Mesh overlay opacity on the video",
    )
    p.add_argument("--device", default="auto", choices=["auto", "cpu", "cuda"])
    return p.parse_args()


def _object_id(video_dir, explicit):
    if explicit:
        return explicit.replace(" ", "_")
    cfg = os.path.join(video_dir, "config.json")
    if os.path.isfile(cfg):
        with open(cfg) as f:
            names = json.load(f).get("object_names") or []
        if names:
            return str(names[0]).replace(" ", "_")
    print("[error] pass --object or put object_names in config.json")
    sys.exit(1)


def _ref_frame(video_dir):
    cfg = os.path.join(video_dir, "config.json")
    if os.path.isfile(cfg):
        with open(cfg) as f:
            return int(json.load(f).get("frame_number", 0))
    return 0


def _anchor_hand(video_dir):
    cfg = os.path.join(video_dir, "config.json")
    if os.path.isfile(cfg):
        with open(cfg) as f:
            h = json.load(f).get("anchor_hand")
        if h in ("left", "right"):
            return h
    return "right"


def _layout_json(video_dir, obj_id, explicit):
    if explicit:
        return explicit
    root = os.path.join(video_dir, "obj_tracking_out", obj_id, "combined_visualization")
    for name in (
        "layout_camera_frame_optimized.json",
        "layout_camera_frame.json",
        "layout.json",
    ):
        path = os.path.join(root, name)
        if os.path.isfile(path):
            return path
    print(f"[error] no layout JSON under {root}")
    sys.exit(1)


def _mesh_path(video_dir, obj_id, explicit):
    if explicit:
        return explicit
    n = _ref_frame(video_dir)
    path = os.path.join(
        video_dir,
        "video_segmentation",
        "masks",
        f"frame_{n:06d}_masks",
        obj_id,
        f"{obj_id}.obj",
    )
    return path


def _rgb_source(video_dir, explicit):
    if explicit:
        return explicit
    clip = os.path.join(video_dir, "clip.mp4")
    frames = os.path.join(video_dir, "all_frames")
    if os.path.isfile(clip):
        return clip
    if os.path.isdir(frames):
        return frames
    print("[error] need clip.mp4 or all_frames/ (or pass --video)")
    sys.exit(1)


def _hand_meshes_npz(video_dir, explicit):
    if explicit:
        return explicit
    clip = os.path.join(video_dir, "clip", "all_hand_meshes.npz")
    if os.path.isfile(clip):
        return clip
    try:
        for name in os.listdir(video_dir):
            path = os.path.join(video_dir, name, "all_hand_meshes.npz")
            if os.path.isfile(path):
                return path
    except FileNotFoundError:
        pass
    return None


def _hand_valid(hand_data, hand, frame_idx):
    key = f"{hand}_valid"
    if key not in hand_data.files:
        return True
    v = np.asarray(hand_data[key])
    if frame_idx >= v.shape[0]:
        return False
    val = np.asarray(v[frame_idx]).reshape(-1)
    if val.size == 0:
        return True
    return bool(val.flat[0])


def _hand_depth_ok(verts, min_median_z=0.15):
    """Skip unused HaWoR side: infiller parks that mesh at the camera (z~0)."""
    z = np.asarray(verts)[:, 2]
    if not np.isfinite(z).all():
        return False
    return float(np.median(z)) >= min_median_z


def main():
    args = parse_args()
    video_dir = os.path.abspath(args.video_dir)
    obj_id = _object_id(video_dir, args.object)
    layout_path = _layout_json(video_dir, obj_id, args.json)
    mesh_path = _mesh_path(video_dir, obj_id, args.mesh)
    video_path = _rgb_source(video_dir, args.video)
    out_path = args.output or os.path.join(
        video_dir, "obj_tracking_out", obj_id, "object_motion.mp4"
    )

    if args.device == "auto":
        device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
    else:
        device = torch.device(args.device)

    print(f"device:  {device}", flush=True)
    print(f"layout:  {layout_path}", flush=True)
    print(f"mesh:    {mesh_path}", flush=True)
    print(f"video:   {video_path}", flush=True)
    print(f"output:  {out_path}", flush=True)

    hand_npz = None if args.hands == "none" else _hand_meshes_npz(video_dir, args.hand_meshes)
    hand_data = None
    hands_to_render = []
    if args.hands != "none":
        if not hand_npz or not os.path.isfile(hand_npz):
            print("[error] HaWoR all_hand_meshes.npz not found (pass --hand-meshes or --hands none)")
            sys.exit(1)
        hand_data = np.load(hand_npz)
        if args.hands == "auto":
            hands_to_render = [_anchor_hand(video_dir)]
        elif args.hands == "both":
            hands_to_render = ["left", "right"]
        else:
            hands_to_render = [args.hands]
        print(f"hands:   {hand_npz}  ({', '.join(hands_to_render)})", flush=True)
    hand_colors = {
        "left": _parse_rgb(args.left_hand_color),
        "right": _parse_rgb(args.right_hand_color),
    }

    with open(layout_path) as f:
        data = json.load(f)
    objects = [o for o in data.get("objects", []) if "frame_idx" in o]
    objects.sort(key=lambda o: o["frame_idx"])
    if not objects:
        print("[error] layout has no frame_idx objects")
        sys.exit(1)

    fallback_intr = None
    for obj in objects:
        if "intrinsics_normalized" in obj:
            fallback_intr = obj["intrinsics_normalized"]
            break
    if fallback_intr is None:
        print("[error] layout has no intrinsics_normalized")
        sys.exit(1)

    is_camera_frame = data.get("frame") == "camera_frame"
    override_scale = None
    tso = data.get("translation_scale_optimization") or {}
    if is_camera_frame and "mesh_scale" in tso:
        override_scale = float(tso["mesh_scale"])
        print(f"using optimized mesh_scale={override_scale:.6f}")

    if not os.path.isfile(mesh_path):
        print(f"[error] mesh not found: {mesh_path}")
        sys.exit(1)
    mesh = load_objs_as_meshes([mesh_path], device=device)
    rgb = _parse_rgb(args.object_color)
    verts_list = mesh.verts_list()
    faces_list = mesh.faces_list()
    n_verts = verts_list[0].shape[0]
    color_t = torch.tensor(rgb, dtype=torch.float32, device=device)
    mesh = Meshes(
        verts=verts_list,
        faces=faces_list,
        textures=TexturesVertex(verts_features=color_t.expand(n_verts, 3).unsqueeze(0)),
    )

    if os.path.isdir(video_path):
        frame_source = DirectoryVideoReader(video_path)
        width, height = frame_source.width, frame_source.height
        cap = None
    else:
        cap = cv2.VideoCapture(video_path)
        width = int(cap.get(cv2.CAP_PROP_FRAME_WIDTH))
        height = int(cap.get(cv2.CAP_PROP_FRAME_HEIGHT))
        frame_source = None

    renderer_cache = {}

    def get_renderer(intr, W, H):
        fx = intr["fx_norm"] * W
        fy = intr["fy_norm"] * H
        cx = intr.get("cx_norm", 0.5) * W
        cy = intr.get("cy_norm", 0.5) * H
        key = (round(fx, 4), round(fy, 4), round(cx, 4), round(cy, 4), W, H)
        if key not in renderer_cache:
            renderer_cache[key] = make_renderer(fx, fy, cx, cy, W, H, device)
        return renderer_cache[key]

    os.makedirs(os.path.dirname(os.path.abspath(out_path)), exist_ok=True)
    fourcc = cv2.VideoWriter_fourcc(*"mp4v")
    tmp_avi = out_path + ".tmp.avi"
    writer = cv2.VideoWriter(tmp_avi, fourcc, args.fps, (width, height))
    if not writer.isOpened():
        print("[error] could not open video writer")
        sys.exit(1)

    last_read_idx = -1
    n_ok = 0
    blend = float(np.clip(args.alpha, 0.0, 1.0))
    for i, obj in enumerate(objects):
        frame_idx = obj["frame_idx"]
        lts = obj["local_to_scene"]
        scale = override_scale if override_scale is not None else lts["scale"][0]
        if is_camera_frame:
            quat = lts["quat_wxyz_camera_frame"]
            translation = list(lts["translation_camera_frame"])
        else:
            quat = lts["new_quat"]
            tx, ty, tz = lts["translation"]
            translation = [tz, tx, ty]

        if frame_source is not None:
            frame_rgb = frame_source.read_frame(frame_idx)
        else:
            frame_rgb, last_read_idx = read_video_frame(cap, frame_idx, last_read_idx)
        if frame_rgb is None:
            print(f"  skip frame {frame_idx} (no RGB)")
            continue

        renderer = get_renderer(obj.get("intrinsics_normalized", fallback_intr), width, height)
        quat_xyzw = np.array([quat[1], quat[2], quat[3], quat[0]])
        rot_matrix = R.from_quat(quat_xyzw).as_matrix()

        obj_verts_cam = transform_to_camera_frame(
            mesh.verts_packed(), rot_matrix, translation, scale, device,
            camera_frame=is_camera_frame,
        )
        meshes_data = [(obj_verts_cam, mesh.faces_packed(), rgb)]
        n_hand_frames = None
        if hand_data is not None:
            n_hand_frames = int(hand_data["right_vertices"].shape[0])
            if frame_idx < n_hand_frames:
                for hand in hands_to_render:
                    verts_np = np.asarray(hand_data[f"{hand}_vertices"][frame_idx])
                    if not _hand_valid(hand_data, hand, frame_idx):
                        continue
                    if not _hand_depth_ok(verts_np):
                        continue
                    verts = torch.tensor(verts_np, dtype=torch.float32, device=device)
                    faces = torch.tensor(
                        hand_data[f"{hand}_faces"], dtype=torch.int64, device=device,
                    )
                    meshes_data.append((verts, faces, hand_colors[hand]))

        if len(meshes_data) > 1:
            render_rgb, render_alpha = render_combined(meshes_data, renderer, device)
        else:
            render_rgb, render_alpha = render_mesh(
                mesh, rot_matrix, translation, scale, renderer, device,
                camera_frame=is_camera_frame,
            )
        render_u8 = (render_rgb * 255).clip(0, 255).astype(np.uint8)
        mask = render_alpha > 0.5
        result = frame_rgb.copy()
        result[mask] = (
            (1.0 - blend) * result[mask].astype(np.float32)
            + blend * render_u8[mask].astype(np.float32)
        ).clip(0, 255).astype(np.uint8)
        cv2.putText(
            result,
            f"frame {frame_idx}",
            (12, 28),
            cv2.FONT_HERSHEY_SIMPLEX,
            0.7,
            (255, 255, 255),
            2,
            cv2.LINE_AA,
        )
        writer.write(cv2.cvtColor(result, cv2.COLOR_RGB2BGR))
        n_ok += 1
        if (i + 1) % 10 == 0 or i == 0:
            print(f"  [{i+1}/{len(objects)}] frame {frame_idx}", flush=True)

    writer.release()
    if cap is not None:
        cap.release()

    # Re-encode to H.264 for browsers / players.
    import subprocess
    cmd = [
        "ffmpeg", "-y", "-i", tmp_avi,
        "-c:v", "libx264", "-pix_fmt", "yuv420p", "-crf", "18",
        out_path,
    ]
    subprocess.run(cmd, check=True)
    os.remove(tmp_avi)
    print(f"Wrote {n_ok} frames -> {out_path}", flush=True)


if __name__ == "__main__":
    main()