| |
| """Check resolution/fps/frame-count alignment across allegro_paired_data video dirs.""" |
|
|
| from __future__ import annotations |
|
|
| import json |
| import pickle |
| import re |
| import subprocess |
| from concurrent.futures import ProcessPoolExecutor, as_completed |
| from pathlib import Path |
|
|
| ROOT = Path("/data/fanliaoyuan/wxy/allegro_paired_data") |
| FFPROBE = "/data/fanliaoyuan/miniconda3/envs/sam3/bin/ffprobe" |
| EXP_W, EXP_H, EXP_FPS = 832, 480, 30.0 |
| TOL_FPS = 0.05 |
|
|
| HUMAN_PAT = re.compile(r"^(.+)-human-view(\d+)\.mp4$") |
|
|
|
|
| def probe_video(path: str) -> dict: |
| cmd = [ |
| FFPROBE, |
| "-v", |
| "error", |
| "-select_streams", |
| "v:0", |
| "-show_entries", |
| "stream=width,height,r_frame_rate,nb_frames,duration", |
| "-of", |
| "json", |
| path, |
| ] |
| out = subprocess.check_output(cmd, text=True) |
| st = json.loads(out)["streams"][0] |
| w, h = int(st["width"]), int(st["height"]) |
| fps_s = st.get("r_frame_rate", "30/1") |
| if "/" in fps_s: |
| n, d = fps_s.split("/") |
| fps = float(n) / float(d) |
| else: |
| fps = float(fps_s) |
| nb = st.get("nb_frames") |
| n_frames = int(nb) if nb and int(nb) > 0 else None |
| dur = st.get("duration") |
| if n_frames is None and dur: |
| n_frames = max(1, int(float(dur) * fps + 0.5)) |
| return {"width": w, "height": h, "fps": fps, "n_frames": n_frames} |
|
|
|
|
| def check_one(args: tuple) -> dict: |
| pack, view = args |
| hv = ROOT / "human_video" / f"{pack}-human-view{view}.mp4" |
| hm = ROOT / "human_video_hand_mask" / f"{pack}-mask-view{view}.mp4" |
| rv = ROOT / "robot_video" / f"{pack}-allegro-view{view}.mp4" |
| pkl = ROOT / "robot_low_dim" / f"{pack}-allegro-view{view}.pkl" |
|
|
| row = {"pack": pack, "view": view, "issues": []} |
| probes = {} |
| for name, path in [("human", hv), ("mask", hm), ("robot", rv)]: |
| try: |
| probes[name] = probe_video(str(path)) |
| except Exception as e: |
| row["issues"].append(f"{name}_probe:{e}") |
|
|
| for name, pr in probes.items(): |
| if pr["width"] != EXP_W or pr["height"] != EXP_H: |
| row["issues"].append(f"{name}_res:{pr['width']}x{pr['height']}") |
| if abs(pr["fps"] - EXP_FPS) > TOL_FPS: |
| row["issues"].append(f"{name}_fps:{pr['fps']:.3f}") |
|
|
| nf = {k: v["n_frames"] for k, v in probes.items()} |
| row["frames"] = nf |
|
|
| if len(set(nf.values())) > 1: |
| row["issues"].append(f"frame_mismatch:{nf}") |
|
|
| try: |
| with open(pkl, "rb") as f: |
| meta = pickle.load(f)["meta"] |
| row["pkl_frames"] = int(meta["num_frames"]) |
| if nf and row["pkl_frames"] not in nf.values(): |
| row["issues"].append(f"pkl_vs_video:pkl={row['pkl_frames']},videos={nf}") |
| except Exception as e: |
| row["issues"].append(f"pkl:{e}") |
|
|
| return row |
|
|
|
|
| def main(): |
| keys = [] |
| for p in (ROOT / "human_video").glob("*.mp4"): |
| m = HUMAN_PAT.match(p.name) |
| if m: |
| keys.append((m.group(1), m.group(2))) |
|
|
| keys.sort() |
| print(f"samples: {len(keys)}") |
|
|
| bad_res = {"human": 0, "mask": 0, "robot": 0} |
| bad_fps = {"human": 0, "mask": 0, "robot": 0} |
| frame_triple_mismatch = 0 |
| pkl_mismatch = 0 |
| other = 0 |
| issue_rows = [] |
| res_counter = {} |
| fps_counter = {} |
|
|
| workers = 24 |
| with ProcessPoolExecutor(max_workers=workers) as ex: |
| futs = [ex.submit(check_one, k) for k in keys] |
| for i, fut in enumerate(as_completed(futs), 1): |
| row = fut.result() |
| if row.get("issues"): |
| issue_rows.append(row) |
| for iss in row["issues"]: |
| if iss.endswith(f":{EXP_W}x{EXP_H}") or "_res:" in iss: |
| pass |
| if "_res:" in iss: |
| for n in ("human", "mask", "robot"): |
| if f"{n}_res:" in iss: |
| bad_res[n] += 1 |
| if "_fps:" in iss: |
| for n in ("human", "mask", "robot"): |
| if f"{n}_fps:" in iss: |
| bad_fps[n] += 1 |
| if iss.startswith("frame_mismatch"): |
| frame_triple_mismatch += 1 |
| if iss.startswith("pkl_vs_video"): |
| pkl_mismatch += 1 |
| if "_probe:" in iss or iss.startswith("pkl:"): |
| other += 1 |
|
|
| |
| if i <= 200 and row.get("frames"): |
| for name, pr_key in [("human", "human"), ("mask", "mask"), ("robot", "robot")]: |
| pass |
|
|
| if i % 500 == 0 or i == len(futs): |
| print(f" [{i}/{len(futs)}] issues so far: {len(issue_rows)}", flush=True) |
|
|
| |
| bad_res = {"human": 0, "mask": 0, "robot": 0} |
| bad_fps = {"human": 0, "mask": 0, "robot": 0} |
| res_vals = set() |
| fps_vals = set() |
| frame_triple_mismatch = 0 |
| pkl_mismatch = 0 |
| for row in issue_rows: |
| for iss in row["issues"]: |
| if "_res:" in iss: |
| for n in ("human", "mask", "robot"): |
| if f"{n}_res:" in iss: |
| bad_res[n] += 1 |
| res_vals.add(iss.split(":", 1)[1]) |
| if "_fps:" in iss: |
| for n in ("human", "mask", "robot"): |
| if f"{n}_fps:" in iss: |
| bad_fps[n] += 1 |
| fps_vals.add(iss.split(":", 1)[1]) |
| if iss.startswith("frame_mismatch"): |
| frame_triple_mismatch += 1 |
| if iss.startswith("pkl_vs_video"): |
| pkl_mismatch += 1 |
|
|
| out = ROOT / "pairing_audit" / "video_alignment_issues.txt" |
| with out.open("w", encoding="utf-8") as f: |
| for row in sorted(issue_rows, key=lambda r: (r["pack"], r["view"])): |
| f.write(f"{row['pack']}\tview{row['view']}\t{'; '.join(row['issues'])}\n") |
|
|
| print("\n=== 分辨率 (期望 832x480) ===") |
| print(f" human 非标准: {bad_res['human']}") |
| print(f" mask 非标准: {bad_res['mask']}") |
| print(f" robot 非标准: {bad_res['robot']}") |
| if res_vals: |
| print(f" 异常分辨率样例: {sorted(res_vals)[:10]}") |
|
|
| print("\n=== 帧率 (期望 30fps) ===") |
| print(f" human 非30fps: {bad_fps['human']}") |
| print(f" mask 非30fps: {bad_fps['mask']}") |
| print(f" robot 非30fps: {bad_fps['robot']}") |
| if fps_vals: |
| print(f" 异常帧率样例: {sorted(fps_vals)[:10]}") |
|
|
| print("\n=== 帧数对齐 ===") |
| print(f" 三视频帧数不一致: {frame_triple_mismatch}") |
| print(f" pkl num_frames 与视频不符: {pkl_mismatch}") |
| print(f" 总问题样本数: {len(issue_rows)} / {len(keys)}") |
| print(f" 详情: {out}") |
|
|
|
|
| if __name__ == "__main__": |
| main() |
|
|
