Implementation:Datajuicer Data juicer VideoSplitBySceneMapper

Knowledge Sources	Datajuicer_Data_juicer
Domains	Data_Processing, Mapping
Last Updated	2026-02-14 16:00 GMT

Overview

Concrete tool for splitting videos into scene clips based on detected scene changes provided by Data-Juicer.

Description

VideoSplitBySceneMapper is a mapper operator that splits videos into individual scene clips based on detected scene changes, using content-aware analysis to identify natural visual boundaries. It uses the scenedetect library with configurable detectors (ContentDetector, ThresholdDetector, or AdaptiveDetector), a threshold parameter, and minimum scene length to identify scene boundaries, then splits the video at those boundaries using FFmpeg, saving individual scene clips and updating the sample's video and text references.

Usage

Use when you need the most semantically meaningful video splitting approach by detecting actual scene changes rather than relying on fixed durations or codec-level keyframes, producing coherent single-scene clips ideal for training.

Code Reference

Source Location

• Repository: Datajuicer_Data_juicer
• File: data_juicer/ops/mapper/video_split_by_scene_mapper.py

Signature

@OPERATORS.register_module("video_split_by_scene_mapper")
class VideoSplitBySceneMapper(Mapper):
    def __init__(self, detector: str = "ContentDetector", threshold: NonNegativeFloat = 27.0, min_scene_len: NonNegativeInt = 15, show_progress: bool = False, save_dir: str = None, save_field: str = None, ffmpeg_extra_args: str = "-movflags frag_keyframe+empty_moov", output_format: str = "path", *args, **kwargs):

Import

from data_juicer.ops.mapper.video_split_by_scene_mapper import VideoSplitBySceneMapper

I/O Contract

Inputs

Name	Type	Required	Description
detector	str	No	Scene detector algorithm: "ContentDetector", "ThresholdDetector", or "AdaptiveDetector" (default: "ContentDetector")
threshold	NonNegativeFloat	No	Threshold passed to the scene detector (default: 27.0)
min_scene_len	NonNegativeInt	No	Minimum length of any scene in frames (default: 15)
show_progress	bool	No	Whether to show progress from scenedetect (default: False)
save_dir	str	No	Directory for generated video files; if not specified, saves alongside input files
save_field	str	No	New field name for generated video paths; if not specified, overwrites original field
ffmpeg_extra_args	str	No	Extra FFmpeg args for splitting (default: "-movflags frag_keyframe+empty_moov")
output_format	str	No	Output format: "path" or "bytes" (default: "path")

Outputs

Name	Type	Description
samples	Dict	Transformed samples with scene-split video clip file paths or bytes

Usage Examples

process:
  - video_split_by_scene_mapper:
      detector: "ContentDetector"
      threshold: 27.0
      min_scene_len: 15

Related Pages

• Environment:Datajuicer_Data_juicer_Python_Runtime_Environment