Video Clipping

Split long videos into shorter clips for downstream processing.

How it Works

NeMo Curator provides two clipping stages: Fixed Stride and TransNetV2 scene-change detection.

• Use Fixed Stride to create uniform segments.

• Use TransNetV2 to cut at visual shot boundaries.

Before You Start

Ensure inputs contain video bytes and basic metadata. The clipping stages require video.source_bytes to be present and metadata with framerate and num_frames.

---

Quickstart

Use either the pipeline stages or the example script flags to create clips.

Pipeline Stage

from nemo_curator.pipeline import Pipeline
from nemo_curator.stages.video.clipping.clip_extraction_stages import (
    FixedStrideExtractorStage,
)
from nemo_curator.stages.video.clipping.video_frame_extraction import (
    VideoFrameExtractionStage,
)
from nemo_curator.stages.video.clipping.transnetv2_extraction import (
    TransNetV2ClipExtractionStage,
)

pipe = Pipeline(name="clipping_examples")

# Fixed Stride
pipe.add_stage(
    FixedStrideExtractorStage(
        clip_len_s=10.0,
        clip_stride_s=10.0,
        min_clip_length_s=2.0,
        limit_clips=0,
    )
)

# TransNetV2 (requires full-video frame extraction first)
pipe.add_stage(VideoFrameExtractionStage(decoder_mode="pynvc", verbose=True))
pipe.add_stage(
    TransNetV2ClipExtractionStage(
        model_dir="/models",
        threshold=0.4,
        min_length_s=2.0,
        max_length_s=10.0,
        max_length_mode="stride",
        crop_s=0.5,
        gpu_memory_gb=10,
        limit_clips=-1,
        verbose=True,
    )
)

pipe.run()

Script Flags

# Fixed stride
python -m nemo_curator.examples.video.video_split_clip_example \
  ... \
  --splitting-algorithm fixed_stride \
  --fixed-stride-split-duration 10.0 \
  --fixed-stride-min-clip-length-s 2.0 \
  --limit-clips 0

# TransNetV2
python -m nemo_curator.examples.video.video_split_clip_example \
  ... \
  --splitting-algorithm transnetv2 \
  --transnetv2-frame-decoder-mode pynvc \
  --transnetv2-threshold 0.4 \
  --transnetv2-min-length-s 2.0 \
  --transnetv2-max-length-s 10.0 \
  --transnetv2-max-length-mode stride \
  --transnetv2-crop-s 0.5 \
  --transnetv2-gpu-memory-gb 10 \
  --limit-clips 0

Clipping Options

Fixed Stride

The FixedStrideExtractorStage steps through the video duration by clip_stride_s, creating spans of length clip_len_s (it truncates the final span at the video end when needed). It filters spans shorter than min_clip_length_s and appends Clip objects identified by source and frame indices.

from nemo_curator.stages.video.clipping.clip_extraction_stages import FixedStrideExtractorStage

stage = FixedStrideExtractorStage(
    clip_len_s=10.0,
    clip_stride_s=10.0,
    min_clip_length_s=2.0,
    limit_clips=0,
)

Tip

If limit_clips > 0 and the Video already has clips, the stage skips processing. It does not cap the number of clips generated within the same run.

TransNetV2 Scene-Change Detection

TransNetV2 is a shot-boundary detection model that identifies transitions between shots. The stage converts those transitions into scenes, applies length/crop rules, and emits clips aligned to scene boundaries.

Using extracted frames of size 27×48×3, the model predicts shot transitions, converts them into scenes, and applies filtering: min_length_s, max_length_s with max_length_mode (“truncate” or “stride”), and optional crop_s at both ends. It creates Clip objects for the resulting spans, then stops after it reaches limit_clips (> 0), and releases frames from memory after processing.

1. Run VideoFrameExtractionStage first to populate video.frame_array.

   from nemo_curator.stages.video.clipping.video_frame_extraction import VideoFrameExtractionStage
   from nemo_curator.stages.video.clipping.transnetv2_extraction import TransNetV2ClipExtractionStage
   
   frame_extractor = VideoFrameExtractionStage(
       decoder_mode="pynvc",  # or "ffmpeg_gpu", "ffmpeg_cpu"
       verbose=True,
   )
   

   Important

   Frames must be (27, 48, 3) per frame; the stage accepts arrays shaped (num_frames, 27, 48, 3) and transposes from (48, 27, 3) automatically.

2. Configure TransNetV2 and run the stage in your pipeline to generate clips from the detected scenes.

   transnet = TransNetV2ClipExtractionStage(
       model_dir="/models",
       threshold=0.4,
       min_length_s=2.0,
       max_length_s=10.0,
       max_length_mode="stride",  # or "truncate"
       crop_s=0.5,
       gpu_memory_gb=10,
       limit_clips=-1,
       verbose=True,
   )