#!/usr/bin/env python3 """Build a LiveWorld four-field CSV from SpatialVID_preped. SpatialVID / LongLive i2v assets: - sampled_clippath.csv -> prompt + clipPath - / -> first-frame image - clip_annotation//longlive/scene_projection_input.mp4 (81 frames) - clip_annotation//longlive/fg_projection_input.mp4 (81 frames) LiveWorld ``condition_source=mp4`` with ``num_frames=81`` indexes projection videos at global indices ``1..81`` (first_frame is separate at index 0). So an 81-frame LongLive control must be padded to 82 frames: lw_proj = [pad_frame] + longlive_proj[0..80] then LiveWorld targets ``1..81`` consume the original 81 LongLive frames. Pad policy: - bg: first frame of the source clip (same as input_image) - fg: black frame (FG at t=0 lives in the RGB image) Writes per stem under ``--output-root//``: first_frame.png, bg_projection.mp4, fg_projection.mp4 and a CSV (paths relative to the CSV file location): name,input_image,text,bg_projection,fg_projection """ from __future__ import annotations import argparse import csv import sys from pathlib import Path from typing import Optional, Tuple import cv2 import numpy as np from tqdm import tqdm def parse_args() -> argparse.Namespace: p = argparse.ArgumentParser(description=__doc__) p.add_argument( "--csv", type=Path, default=Path( "/home/zhangzhiyuan/VAM_projs/realbench_comp/" "SpatialVID_preped/sampled_clippath.csv" ), ) p.add_argument( "--video-root", type=Path, default=Path( "/home/zhangzhiyuan/VAM_projs/realbench_comp/SpatialVID_preped" ), help="Directory containing mp4s", ) p.add_argument( "--annotation-root", type=Path, default=Path( "/home/zhangzhiyuan/VAM_projs/realbench_comp/" "SpatialVID_preped/clip_annotation" ), ) p.add_argument( "--output-root", type=Path, default=Path(__file__).resolve().parents[1] / "liveworld_i2v_inputs", help="Per-case first_frame + padded 82-frame projections " "(default: LiveWorld_comp/liveworld_i2v_inputs)", ) p.add_argument( "--out-csv", type=Path, default=Path(__file__).resolve().parents[1] / "liveworld_i2v_cases.csv", help="Output CSV path (default: LiveWorld_comp/liveworld_i2v_cases.csv)", ) p.add_argument("--fps", type=float, default=16.0, help="FPS for written projection mp4s (LiveWorld default 16)") p.add_argument("--target-hw", type=int, nargs=2, default=[480, 832], metavar=("H", "W"), help="Resize first_frame / projections to this HxW") p.add_argument("--overwrite", action="store_true") p.add_argument("--limit", type=int, default=None) p.add_argument("--stems", nargs="*", default=None, help="Only process these stems (default: all csv rows)") return p.parse_args() def read_video_bgr(path: Path) -> list: cap = cv2.VideoCapture(str(path)) if not cap.isOpened(): raise FileNotFoundError(f"cannot open video: {path}") frames = [] while True: ok, frame = cap.read() if not ok: break frames.append(frame) cap.release() if not frames: raise RuntimeError(f"empty video: {path}") return frames def write_video_bgr(path: Path, frames: list, fps: float) -> None: path.parent.mkdir(parents=True, exist_ok=True) h, w = frames[0].shape[:2] writer = cv2.VideoWriter( str(path), cv2.VideoWriter_fourcc(*"mp4v"), float(fps), (w, h), ) if not writer.isOpened(): raise RuntimeError(f"failed to open VideoWriter: {path}") for f in frames: writer.write(f) writer.release() def resize_bgr(frame: np.ndarray, target_hw: Tuple[int, int]) -> np.ndarray: H, W = target_hw if frame.shape[0] == H and frame.shape[1] == W: return frame return cv2.resize(frame, (W, H), interpolation=cv2.INTER_LINEAR) def pad_projection_to_82( frames_81: list, pad_frame: np.ndarray, target_hw: Tuple[int, int], ) -> list: """Prepend pad_frame so LiveWorld indices 1..81 map to original 0..80.""" if len(frames_81) != 81: raise ValueError(f"expected 81 frames, got {len(frames_81)}") pad = resize_bgr(pad_frame, target_hw) body = [resize_bgr(f, target_hw) for f in frames_81] return [pad] + body def process_one( stem: str, prompt: str, video_path: Path, annotation_root: Path, output_root: Path, fps: float, target_hw: Tuple[int, int], overwrite: bool, ) -> Optional[dict]: ll_dir = annotation_root / stem / "longlive" bg_src = ll_dir / "scene_projection_input.mp4" fg_src = ll_dir / "fg_projection_input.mp4" if not bg_src.exists() or not fg_src.exists(): print(f"[skip] missing projections: {stem}", file=sys.stderr) return None if not video_path.exists(): print(f"[skip] missing video: {video_path}", file=sys.stderr) return None out_dir = output_root / stem first_path = out_dir / "first_frame.png" bg_out = out_dir / "bg_projection.mp4" fg_out = out_dir / "fg_projection.mp4" if ( not overwrite and first_path.exists() and bg_out.exists() and fg_out.exists() ): return { "name": stem, "input_image": first_path, "text": prompt, "bg_projection": bg_out, "fg_projection": fg_out, } src_frames = read_video_bgr(video_path) first_bgr = resize_bgr(src_frames[0], target_hw) out_dir.mkdir(parents=True, exist_ok=True) cv2.imwrite(str(first_path), first_bgr) bg_81 = read_video_bgr(bg_src) fg_81 = read_video_bgr(fg_src) if len(bg_81) != 81 or len(fg_81) != 81: print( f"[skip] {stem}: expected 81-frame projections, " f"got bg={len(bg_81)} fg={len(fg_81)}", file=sys.stderr, ) return None H, W = target_hw black = np.zeros((H, W, 3), dtype=np.uint8) bg_82 = pad_projection_to_82(bg_81, first_bgr, target_hw) fg_82 = pad_projection_to_82(fg_81, black, target_hw) write_video_bgr(bg_out, bg_82, fps) write_video_bgr(fg_out, fg_82, fps) return { "name": stem, "input_image": first_path, "text": prompt, "bg_projection": bg_out, "fg_projection": fg_out, } def main() -> None: args = parse_args() target_hw = (int(args.target_hw[0]), int(args.target_hw[1])) wanted = set(args.stems) if args.stems else None with args.csv.open("r", encoding="utf-8-sig", newline="") as f: rows = list(csv.DictReader(f)) if not rows: raise SystemExit(f"empty csv: {args.csv}") selected = [] for r in rows: clip = (r.get("clipPath") or "").strip() if not clip: continue stem = Path(clip).stem if wanted is not None and stem not in wanted: continue selected.append((stem, (r.get("prompt") or "").strip(), clip)) if args.limit is not None: selected = selected[: args.limit] args.output_root.mkdir(parents=True, exist_ok=True) out_rows = [] n_fail = 0 for stem, prompt, clip in tqdm(selected, desc="prepare liveworld i2v"): if not prompt: print(f"[skip] empty prompt: {stem}", file=sys.stderr) n_fail += 1 continue video_path = args.video_root / clip try: row = process_one( stem=stem, prompt=prompt, video_path=video_path, annotation_root=args.annotation_root, output_root=args.output_root, fps=args.fps, target_hw=target_hw, overwrite=args.overwrite, ) except Exception as e: # noqa: BLE001 n_fail += 1 print(f"[FAIL] {stem}: {e}", file=sys.stderr) continue if row is None: n_fail += 1 continue out_rows.append(row) args.out_csv.parent.mkdir(parents=True, exist_ok=True) csv_base = args.out_csv.resolve().parent def _rel(p: Path) -> str: try: return str(p.resolve().relative_to(csv_base)) except ValueError: return str(p.resolve()) csv_rows = [ { "name": r["name"], "input_image": _rel(Path(r["input_image"])), "text": r["text"], "bg_projection": _rel(Path(r["bg_projection"])), "fg_projection": _rel(Path(r["fg_projection"])), } for r in out_rows ] with args.out_csv.open("w", encoding="utf-8", newline="") as f: writer = csv.DictWriter( f, fieldnames=[ "name", "input_image", "text", "bg_projection", "fg_projection", ], quoting=csv.QUOTE_MINIMAL, ) writer.writeheader() writer.writerows(csv_rows) print( f"[done] wrote {len(csv_rows)} rows -> {args.out_csv} " f"(fail/skip={n_fail}, inputs={args.output_root})" ) if __name__ == "__main__": main()