Run PySceneDetect to get raw scene boundaries (in seconds). Returns: (scene_pairs_in_seconds, fps) Notes: - Uses Content

Discover image files for a given clip: <figure_root>/<stem>/<clip_id>/img/*.jpg

Discover image files for a given clip: <figure_root>/<stem>/<clip_id>/img/*.jpg Returns all found images.

Discover image files for a given clip: <figure_root>/<stem>/<clip_id>/img/*.jpg

save generated image from API response

Extract content between <answer> and </answer> tags.

Extract content between <answer> and </answer> tags.

Extract CLIP embeddings for a list of images. Args: images: List of images in various formats

Extract CLIP embeddings for a list of images. Args: images: List of images in various formats

Top-level picklable worker that extracts frames for a single clip within a video. Returns a dict containing per-clip results.

Extract multi-modal token embeddings and sample features from Qwen2-VL. Returns: Tuple of (token_features_list,

Extract content between <think> and </think> tags.

Extract content between <think> and </think> tags.

Extract original video name from clip ID. Example: "trailer_0" -> "trailer", "video_name_5" -> "video_name"

Filter a single multi-modal sample. Returns True if keep, False if filter out.

Find duplicate images based on embedding similarity. Args: embeddings: Image embeddings of shape (n_images, embe

把你给的 video_clips（list，每元素含 {success, error, clips:[...] }） 展平成 DataFrame（每条 clip 一行）。

Helper to generate images for a batch of prompts on a single pipeline.

Normalize scenes list: - Prefer dict format from `video_scene['scenes']` - If cell is malformed or empty, return []

获取视频分辨率 Args: video_path: 视频文件路径 Returns: tuple: (width, height)，如果失败返回 (-1, -1

Load NSFW detection model.

计算两个字符串的最长公共前缀长度

Build per-row clip entries from scene boundaries and video_info. Returns {"success": bool, "error": Optional[str], "clips": [clip dicts]}

直接使用 self.pool 分发任务，不再创建Pool。

解析文本中的多模态token，并返回它们的计数和去除token后的文本。

解析文本中的多模态token，并返回它们的计数和去除token后的文本。

解析文本中的多模态token，并返回它们的计数和去除token后的文本。

解析文本中的多模态token，并返回它们的计数和去除token后的文本。

解析文本中的多模态token，并返回它们的计数和去除token后的文本。

Accepts "640*360" or (640, 360) or None. Returns (w, h) or None.

V_i = (1/3) * V_i^Inf + (1/3) * (V_i^Uni + V_i^Rep)

Prepare messages format for Qwen2-VL.

Prepare regex patterns for text filtering.

Print summary of final results.

Print summary of final results.

Probe a single video path with the chosen backend. This is exception-safe and returns a structured VideoInfo. Notes: - If 'path'

Probe basic video info using PyAV. Returns: (num_frames, height, width, fps, duration_sec) Notes: PyAV can provide dura

Probe basic video info using OpenCV. Returns: (num_frames, height, width, fps) Raises: ValueError if the video cannot be

Probe basic video info using TorchVision's read_video. Returns: (num_frames, height, width, fps) Notes: Assumes your lo

Process CoT QA responses to extract answers and think chains.

Process CoT QA responses to extract answers and think chains.

Compute pixel count safely.

Resolve image path. If it's a relative path, resolve it to an absolute path relative to base_dir. If it's an absolute path, return it

Sample up to `load_num` frames from the whole video (evenly spaced). Used when we don't have extracted frames on disk (e.g. video_clips_key=N

Sample up to `load_num` frames from the whole video (evenly spaced) and return PIL Images. Used when we don't have extracted frames on disk (

Sample up to `load_num` frames from the whole video (evenly spaced). Returns RGB uint8 tensors with shape (C,H,W). Used when we don't hav

串行处理器：在一个循环中处理所有音频。

Safely set the first user message's 'value' in a conversation. Expected format: a list of messages (dicts), where the first item repr

Safely set the first user message's 'value' in a conversation. Expected format: a list of messages (dicts), where the first item repr

Safely set the first user message's 'value' in a conversation.

Split video by fixed interval, similar to step1_split_videos.py logic. Args: video_duration: Total video duration in seconds

Apply head/tail trimming and split long scenes into chunks within [min, max]. Returns a list of (start_sec, end_sec).

Apply head/tail trimming and split long scenes with overlap strategy. Similar to clip_caption.py logic: - If scene duration > max_seconds

Analyze multiple images in a single request with labels. :param image_paths: List of image file paths. :param image_labels:

Returns the names of available CLIP models

Calculate reward based on question type and model output. Args: gt_solution: Ground truth solution (with <answer> tags)

Perform a chat completion using a single image and a text prompt. :param image_path: Path to the image file. :param text_pro

Combine multiple images into one and return the combined image as a Base64 string. :param image_paths: List of image file paths.

Calculate average ROUGE F-measure.