scripts/visualize_robotrack_dataset.py

#!/usr/bin/env python3
"""Render RoboTrack point annotations on top of their source videos.

Expected layout:

    DATASET_ROOT/
      clip_id/
        video.mp4
        point_tracks.npz

Each NPZ must contain:

    trajs_2d:     (frames, tracks, 2) pixel coordinates
    visibility:   (frames, tracks) visibility scores
    query_frames: (tracks,) first/query frame for each track

The default output is ``point_track_vis.mp4`` in each clip directory. Existing
outputs are skipped unless ``--overwrite`` is supplied, so interrupted runs can
be resumed safely.
"""

from __future__ import annotations

import argparse
import colorsys
import os
from concurrent.futures import ProcessPoolExecutor, as_completed
from pathlib import Path
import shutil
import subprocess
import sys

import cv2
import numpy as np


DEFAULT_FFMPEG_CANDIDATES = (
    "/gpfs/projects/raivn/yunbos/.conda/envs/cotracker-perception/bin/ffmpeg",
)


def parse_args() -> argparse.Namespace:
    parser = argparse.ArgumentParser(description=__doc__)
    parser.add_argument("dataset_root", type=Path, help="RoboTrack dataset directory")
    parser.add_argument("--video-name", default="video.mp4")
    parser.add_argument("--tracks-name", default="point_tracks.npz")
    parser.add_argument("--output-name", default="point_track_vis.mp4")
    parser.add_argument(
        "--trail-seconds",
        type=float,
        default=1.0,
        help="Length of the visible motion trail (default: 1.0)",
    )
    parser.add_argument(
        "--visibility-threshold",
        type=float,
        default=0.5,
        help="Scores above this value are drawn as visible (default: 0.5)",
    )
    parser.add_argument(
        "--crf",
        type=int,
        default=20,
        help="H.264 quality: lower is better/larger (default: 20)",
    )
    parser.add_argument(
        "--preset",
        default="veryfast",
        help="libx264 encoding preset (default: veryfast)",
    )
    parser.add_argument(
        "--workers",
        type=int,
        default=min(4, os.cpu_count() or 1),
        help="Parallel clips to render (default: up to 4)",
    )
    parser.add_argument(
        "--limit",
        type=int,
        help="Render only the first N clips (useful for testing)",
    )
    parser.add_argument("--overwrite", action="store_true")
    parser.add_argument(
        "--ffmpeg",
        type=Path,
        help="Path to ffmpeg; otherwise resolve it automatically",
    )
    return parser.parse_args()


def find_ffmpeg(explicit_path: Path | None) -> str:
    if explicit_path is not None:
        if not explicit_path.is_file():
            raise FileNotFoundError(f"ffmpeg does not exist: {explicit_path}")
        return str(explicit_path.resolve())

    on_path = shutil.which("ffmpeg")
    if on_path:
        return on_path

    for candidate in DEFAULT_FFMPEG_CANDIDATES:
        if Path(candidate).is_file():
            return candidate

    raise FileNotFoundError("Could not find ffmpeg; pass its path with --ffmpeg")


def track_colors(count: int) -> list[tuple[int, int, int]]:
    """Return visually separated, stable BGR colors."""
    colors = []
    golden_ratio = 0.618033988749895
    for index in range(count):
        hue = (0.07 + index * golden_ratio) % 1.0
        red, green, blue = colorsys.hsv_to_rgb(hue, 0.88, 1.0)
        colors.append((round(blue * 255), round(green * 255), round(red * 255)))
    return colors


def validate_tracks(
    npz_path: Path,
) -> tuple[np.ndarray, np.ndarray, np.ndarray]:
    with np.load(npz_path) as data:
        required = {"trajs_2d", "visibility", "query_frames"}
        missing = required.difference(data.files)
        if missing:
            raise ValueError(f"missing NPZ arrays: {', '.join(sorted(missing))}")
        trajectories = np.asarray(data["trajs_2d"], dtype=np.float32)
        visibility = np.asarray(data["visibility"], dtype=np.float32)
        query_frames = np.asarray(data["query_frames"], dtype=np.int64)

    if trajectories.ndim != 3 or trajectories.shape[-1] != 2:
        raise ValueError(f"trajs_2d must have shape (T, N, 2), got {trajectories.shape}")
    if visibility.shape != trajectories.shape[:2]:
        raise ValueError(
            f"visibility shape {visibility.shape} does not match {trajectories.shape[:2]}"
        )
    if query_frames.shape != (trajectories.shape[1],):
        raise ValueError(
            f"query_frames shape {query_frames.shape} does not match "
            f"({trajectories.shape[1]},)"
        )
    if np.any(query_frames < 0) or np.any(query_frames >= trajectories.shape[0]):
        raise ValueError("query_frames contains an index outside the video")
    return trajectories, visibility, query_frames


def visible_segments(
    points: np.ndarray, visible: np.ndarray
) -> list[np.ndarray]:
    """Split a short trajectory window into contiguous visible polylines."""
    segments: list[np.ndarray] = []
    start = None
    for index, is_visible in enumerate(visible):
        if is_visible and np.isfinite(points[index]).all():
            if start is None:
                start = index
        elif start is not None:
            if index - start >= 2:
                segments.append(points[start:index])
            start = None
    if start is not None and len(points) - start >= 2:
        segments.append(points[start:])
    return segments


def outlined_text(
    frame: np.ndarray,
    text: str,
    origin: tuple[int, int],
    font_scale: float,
    color: tuple[int, int, int],
    thickness: int,
) -> None:
    cv2.putText(
        frame,
        text,
        origin,
        cv2.FONT_HERSHEY_SIMPLEX,
        font_scale,
        (0, 0, 0),
        thickness + 3,
        cv2.LINE_AA,
    )
    cv2.putText(
        frame,
        text,
        origin,
        cv2.FONT_HERSHEY_SIMPLEX,
        font_scale,
        color,
        thickness,
        cv2.LINE_AA,
    )


def fit_text_to_width(
    text: str,
    max_width: int,
    font_scale: float,
    thickness: int,
) -> str:
    """Elide the middle of text while preserving its identifying suffix."""
    def width(candidate: str) -> int:
        size, _ = cv2.getTextSize(
            candidate, cv2.FONT_HERSHEY_SIMPLEX, font_scale, thickness
        )
        return size[0]

    if width(text) <= max_width:
        return text
    for keep in range(len(text) - 1, 5, -1):
        prefix_length = (keep + 1) // 2
        suffix_length = keep // 2
        candidate = f"{text[:prefix_length]}...{text[-suffix_length:]}"
        if width(candidate) <= max_width:
            return candidate
    return "..."


def draw_overlay(
    frame: np.ndarray,
    frame_index: int,
    trajectories: np.ndarray,
    visibility: np.ndarray,
    query_frames: np.ndarray,
    colors: list[tuple[int, int, int]],
    trail_frames: int,
    visibility_threshold: float,
    clip_id: str,
) -> np.ndarray:
    height, width = frame.shape[:2]
    num_frames, num_tracks = trajectories.shape[:2]
    visible_now = visibility[frame_index] > visibility_threshold
    visible_now &= query_frames <= frame_index

    point_radius = max(4, round(min(width, height) / 120))
    point_outline = max(2, round(point_radius / 3))
    trail_width = max(2, round(point_radius / 2))
    font_scale = min(1.0, max(0.5, min(width, height) / 900))
    font_thickness = max(1, round(font_scale * 2))

    trail_layer = frame.copy()
    first_trail_frame = max(0, frame_index - trail_frames)
    for track_index in range(num_tracks):
        first = max(first_trail_frame, int(query_frames[track_index]))
        points = trajectories[first : frame_index + 1, track_index]
        visible = visibility[first : frame_index + 1, track_index] > visibility_threshold
        for segment in visible_segments(points, visible):
            rounded = np.rint(segment).astype(np.int32).reshape((-1, 1, 2))
            cv2.polylines(
                trail_layer,
                [rounded],
                isClosed=False,
                color=colors[track_index],
                thickness=trail_width,
                lineType=cv2.LINE_AA,
            )
    cv2.addWeighted(trail_layer, 0.72, frame, 0.28, 0.0, dst=frame)

    for track_index in range(num_tracks):
        if not visible_now[track_index]:
            continue
        point = trajectories[frame_index, track_index]
        if not np.isfinite(point).all():
            continue
        x, y = np.rint(point).astype(int)
        # Coordinates just outside the image can occur in hand-authored tracks.
        # Clipping keeps the renderer robust while still placing a marker at the edge.
        x = int(np.clip(x, 0, width - 1))
        y = int(np.clip(y, 0, height - 1))

        if frame_index == int(query_frames[track_index]):
            cv2.circle(
                frame,
                (x, y),
                point_radius + point_outline + 3,
                (255, 255, 255),
                point_outline,
                cv2.LINE_AA,
            )
        cv2.circle(
            frame,
            (x, y),
            point_radius + point_outline,
            (0, 0, 0),
            -1,
            cv2.LINE_AA,
        )
        cv2.circle(
            frame,
            (x, y),
            point_radius,
            colors[track_index],
            -1,
            cv2.LINE_AA,
        )
        label_x = min(width - 1, x + point_radius + 4)
        label_y = int(np.clip(y - point_radius - 2, 14, height - 2))
        outlined_text(
            frame,
            str(track_index),
            (label_x, label_y),
            font_scale * 0.78,
            colors[track_index],
            font_thickness,
        )

    active_count = int(np.count_nonzero(query_frames <= frame_index))
    clip_line = fit_text_to_width(
        f"clip: {clip_id}", width - 20, font_scale, font_thickness
    )
    stats_line = (
        f"frame {frame_index + 1}/{num_frames}   "
        f"visible {int(np.count_nonzero(visible_now))}/{active_count}   "
        f"tracks {num_tracks}"
    )
    clip_size, baseline = cv2.getTextSize(
        clip_line, cv2.FONT_HERSHEY_SIMPLEX, font_scale, font_thickness
    )
    stats_size, _ = cv2.getTextSize(
        stats_line, cv2.FONT_HERSHEY_SIMPLEX, font_scale, font_thickness
    )
    line_gap = max(5, round(font_scale * 6))
    header_height = clip_size[1] + stats_size[1] + baseline + line_gap + 18
    header_width = min(width, max(clip_size[0], stats_size[0]) + 20)
    header_layer = frame.copy()
    cv2.rectangle(header_layer, (0, 0), (header_width, header_height), (0, 0, 0), -1)
    cv2.addWeighted(header_layer, 0.62, frame, 0.38, 0.0, dst=frame)
    cv2.putText(
        frame,
        clip_line,
        (10, clip_size[1] + 7),
        cv2.FONT_HERSHEY_SIMPLEX,
        font_scale,
        (255, 255, 255),
        font_thickness,
        cv2.LINE_AA,
    )
    cv2.putText(
        frame,
        stats_line,
        (10, clip_size[1] + line_gap + stats_size[1] + 7),
        cv2.FONT_HERSHEY_SIMPLEX,
        font_scale,
        (255, 255, 255),
        font_thickness,
        cv2.LINE_AA,
    )
    return frame


def render_clip(
    clip_dir_string: str,
    video_name: str,
    tracks_name: str,
    output_name: str,
    trail_seconds: float,
    visibility_threshold: float,
    crf: int,
    preset: str,
    ffmpeg: str,
    overwrite: bool,
) -> tuple[str, str, str]:
    clip_dir = Path(clip_dir_string)
    video_path = clip_dir / video_name
    tracks_path = clip_dir / tracks_name
    output_path = clip_dir / output_name
    clip_id = clip_dir.name

    if output_path.exists() and not overwrite:
        return clip_id, "skipped", "already exists"

    trajectories, visibility, query_frames = validate_tracks(tracks_path)
    capture = cv2.VideoCapture(str(video_path))
    if not capture.isOpened():
        raise RuntimeError(f"could not open video: {video_path}")

    width = int(capture.get(cv2.CAP_PROP_FRAME_WIDTH))
    height = int(capture.get(cv2.CAP_PROP_FRAME_HEIGHT))
    fps = float(capture.get(cv2.CAP_PROP_FPS))
    reported_frames = int(capture.get(cv2.CAP_PROP_FRAME_COUNT))
    if width <= 0 or height <= 0 or fps <= 0:
        capture.release()
        raise ValueError(f"invalid video metadata: {width}x{height} at {fps} fps")
    if reported_frames > 0 and reported_frames != trajectories.shape[0]:
        capture.release()
        raise ValueError(
            f"video reports {reported_frames} frames but tracks have "
            f"{trajectories.shape[0]}"
        )

    temporary_path = output_path.with_name(
        f".{output_path.stem}.tmp-{os.getpid()}{output_path.suffix}"
    )
    command = [
        ffmpeg,
        "-hide_banner",
        "-loglevel",
        "error",
        "-y",
        "-f",
        "rawvideo",
        "-pixel_format",
        "bgr24",
        "-video_size",
        f"{width}x{height}",
        "-framerate",
        f"{fps:.8f}",
        "-i",
        "-",
        "-an",
        "-vf",
        "pad=ceil(iw/2)*2:ceil(ih/2)*2",
        "-c:v",
        "libx264",
        "-preset",
        preset,
        "-crf",
        str(crf),
        "-pix_fmt",
        "yuv420p",
        "-movflags",
        "+faststart",
        str(temporary_path),
    ]

    encoder = subprocess.Popen(
        command,
        stdin=subprocess.PIPE,
        stdout=subprocess.DEVNULL,
        stderr=subprocess.PIPE,
    )
    frames_written = 0
    colors = track_colors(trajectories.shape[1])
    trail_frames = max(0, round(trail_seconds * fps))
    failure: Exception | None = None
    try:
        assert encoder.stdin is not None
        for frame_index in range(trajectories.shape[0]):
            ok, frame = capture.read()
            if not ok:
                raise RuntimeError(
                    f"video ended after {frames_written}/{trajectories.shape[0]} frames"
                )
            draw_overlay(
                frame,
                frame_index,
                trajectories,
                visibility,
                query_frames,
                colors,
                trail_frames,
                visibility_threshold,
                clip_id,
            )
            encoder.stdin.write(frame.tobytes())
            frames_written += 1
    except Exception as error:
        failure = error
    finally:
        capture.release()
        if encoder.stdin is not None:
            try:
                encoder.stdin.close()
            except BrokenPipeError:
                pass

    assert encoder.stderr is not None
    encoder_error = encoder.stderr.read().decode("utf-8", errors="replace").strip()
    return_code = encoder.wait()
    if failure is not None or return_code != 0:
        temporary_path.unlink(missing_ok=True)
        details = str(failure) if failure is not None else ""
        if encoder_error:
            details = f"{details}; ffmpeg: {encoder_error}".strip("; ")
        raise RuntimeError(details or f"ffmpeg exited with status {return_code}")

    if frames_written != trajectories.shape[0]:
        temporary_path.unlink(missing_ok=True)
        raise RuntimeError(
            f"wrote {frames_written} frames, expected {trajectories.shape[0]}"
        )
    os.replace(temporary_path, output_path)
    return clip_id, "rendered", f"{frames_written} frames"


def main() -> int:
    args = parse_args()
    dataset_root = args.dataset_root.resolve()
    if not dataset_root.is_dir():
        print(f"error: dataset root does not exist: {dataset_root}", file=sys.stderr)
        return 2
    if args.workers < 1:
        print("error: --workers must be at least 1", file=sys.stderr)
        return 2
    if args.trail_seconds < 0:
        print("error: --trail-seconds cannot be negative", file=sys.stderr)
        return 2
    if not 0 <= args.crf <= 51:
        print("error: --crf must be between 0 and 51", file=sys.stderr)
        return 2

    try:
        ffmpeg = find_ffmpeg(args.ffmpeg)
    except FileNotFoundError as error:
        print(f"error: {error}", file=sys.stderr)
        return 2

    clip_dirs = sorted(
        path
        for path in dataset_root.iterdir()
        if path.is_dir()
        and (path / args.video_name).is_file()
        and (path / args.tracks_name).is_file()
    )
    if args.limit is not None:
        if args.limit < 0:
            print("error: --limit cannot be negative", file=sys.stderr)
            return 2
        clip_dirs = clip_dirs[: args.limit]
    if not clip_dirs:
        print("No matching clip directories found.")
        return 0

    print(
        f"Rendering {len(clip_dirs)} clips from {dataset_root} with "
        f"{args.workers} worker(s)",
        flush=True,
    )
    print(f"ffmpeg: {ffmpeg}", flush=True)

    rendered = 0
    skipped = 0
    failures: list[tuple[str, str]] = []
    common_args = (
        args.video_name,
        args.tracks_name,
        args.output_name,
        args.trail_seconds,
        args.visibility_threshold,
        args.crf,
        args.preset,
        ffmpeg,
        args.overwrite,
    )
    with ProcessPoolExecutor(max_workers=args.workers) as executor:
        future_to_clip = {
            executor.submit(render_clip, str(clip_dir), *common_args): clip_dir.name
            for clip_dir in clip_dirs
        }
        for completed, future in enumerate(as_completed(future_to_clip), start=1):
            clip_id = future_to_clip[future]
            try:
                _, status, detail = future.result()
                if status == "rendered":
                    rendered += 1
                else:
                    skipped += 1
                print(
                    f"[{completed:>3}/{len(clip_dirs)}] {status:8} {clip_id} "
                    f"({detail})",
                    flush=True,
                )
            except Exception as error:
                failures.append((clip_id, str(error)))
                print(
                    f"[{completed:>3}/{len(clip_dirs)}] FAILED   {clip_id}: {error}",
                    file=sys.stderr,
                    flush=True,
                )

    print(
        f"Done: {rendered} rendered, {skipped} skipped, {len(failures)} failed.",
        flush=True,
    )
    if failures:
        print("Failures:", file=sys.stderr)
        for clip_id, error in failures:
            print(f"  {clip_id}: {error}", file=sys.stderr)
        return 1
    return 0


if __name__ == "__main__":
    raise SystemExit(main())