analyze.py

import pandas as pd
data = pd.read_csv("OpenHumanVid_part_001.csv")
import re

# Patterns that indicate multiple characters
multi_person_pattern = re.compile(
    r'\b('
    r'two|three|four|five|six|seven|eight|nine|ten'
    r'|multiple|several|some|a few|a couple|a pair|a group|a crowd|a team|a gang|a duo'
    r'|both|together|each other|one another'
    r'|2|3|4|5|6|7|8|9'
    r')\b',
    re.IGNORECASE
)

single_char_mask = ~data['caption'].str.contains(multi_person_pattern)
single_char_data = data[single_char_mask].reset_index(drop=True)

print(f"Original: {len(data):,} rows")
print(f"Single character: {len(single_char_data):,} rows ({len(single_char_data)/len(data)*100:.1f}%)")
single_char_data

action_patterns = {
    'talking':  r'\b(talk(?:ing|s)?|speak(?:ing|s)?|chat(?:ting|s)?|convers(?:ing|ation)|address(?:ing)?|narrat(?:ing|es?)|monologue|lectur(?:ing|es?))\b',
    'running':  r'\b(runn(?:ing|s)?|jog(?:ging|s)?|sprint(?:ing|s)?|dash(?:ing|es)?)\b',
    'walking':  r'\b(walk(?:ing|s)?|stroll(?:ing|s)?|march(?:ing|es)?|saunter(?:ing)?|striding|strides?)\b',
    'dancing':  r'\b(danc(?:ing|es?|er)?|ballet|choreograph(?:y|ing)|twirl(?:ing)?|sway(?:ing)?)\b',
    'fighting': r'\b(fight(?:ing|s)?|punch(?:ing|es)?|kick(?:ing|s)?|spar(?:ring)?|combat(?:ting)?|battle|martial arts?|box(?:ing)?)\b',
    'singing':  r'\b(sing(?:ing|s)?|sang|song|vocal(?:izing)?|perform(?:ing)? (?:a )?song)\b',
    'sitting':  r'\b(sit(?:ting|s)?|seated|sat down?|lounging|resting)\b',
    'standing': r'\b(stand(?:ing|s)?|stood|upright|posing|poses?)\b',
    'exercising': r'\b(exercis(?:ing|es?)|workout|work(?:ing)? out|gym|yoga|stretch(?:ing)?|lift(?:ing)? weights?|train(?:ing)?)\b',
    'eating':   r'\b(eat(?:ing|s)?|ate|dining|chew(?:ing)?|drinking|drink(?:ing|s)?)\b',
}

samples = {}
for action, pattern in action_patterns.items():
    matched = single_char_data[single_char_data['caption'].str.contains(pattern, case=False, regex=True)]
    n = min(100, len(matched))
    samples[action] = matched.sample(n=n, random_state=42) if n > 0 else matched
    print(f"{action:12s}: {len(matched):5,} matches  →  {n} sampled")

action_samples = pd.concat(samples.values(), keys=samples.keys()).rename_axis(['action', 'idx']).reset_index(level=0)
action_samples = action_samples.reset_index(drop=True)
print(f"\nTotal samples: {len(action_samples)}")
action_samples[['action', 'clip_id', 'path', 'caption']]


import shutil
from pathlib import Path

# CSV paths are "clips/part_001/..." but files live under BASE_DIR/part_001/...
BASE_DIR = Path("/mnt/data/xinyuy/datasets/OpenHumanVid")
OUT_DIR  = BASE_DIR / "selected_2"

copied, missing = 0, 0

for _, row in action_samples.iterrows():
    # Strip the leading "clips/" prefix that doesn't exist on disk
    rel_path = row["path"].removeprefix("clips/")
    src = BASE_DIR / (rel_path + ".mp4")
    dst_dir = OUT_DIR / row["action"]
    dst_dir.mkdir(parents=True, exist_ok=True)
    dst = dst_dir / src.name

    if src.exists():
        shutil.copy2(src, dst)
        copied += 1
    else:
        print(f"[missing] {src}")
        missing += 1

print(f"\nDone — copied {copied} files, {missing} missing.")
print(f"Output: {OUT_DIR}")