ShotDetector.FastYuv 3.30.0

ShotDetector.FastYuv

Optional fast path for ShotDetector: with it, ffmpeg sends raw yuv420p (or yuv422p) and ShotDetector converts only the ~7% of pixels its resize reads, instead of ffmpeg converting every full frame to BGR. Results are identical, and large video is faster (the current numbers: ShotDetector's weekly benchmarks).

using ShotDetector;
using ShotDetector.FastYuv;

var result = ShotDetection.Detect("video.mp4", new DetectionOptions { Yuv420Converter = new SwscaleYuv420() });

It also has Yadif, a port of FFmpeg's yadif deinterlacer, row by row: with in-process decoding and deinterlacing on, only the rows the resize reads are deinterlaced (identical results, and only those rows' work). It is a FrameReader IRowDeinterlacer, so a FrameDecoder takes it too (FrameDecoderOptions.RowDeinterlacer).

var result = ShotDetection.Detect("broadcast.mxf", new DetectionOptions { Deinterlace = DeinterlaceMode.Auto, Yadif = new Yadif() });

The converter is a bit-exact port of FFmpeg libswscale's yuv420p → BGR24 conversion as x86 runs it (verified for all 2^24 Y/U/V values), so this package is licensed under the GNU LGPL 2.1 or later (COPYING.LGPL). The ShotDetector package itself is MIT. It converts yuv422p (and yuvj422p) too: swscale runs the same converter for it, with a chroma row for every row (IYuv422Converter, which SwscaleYuv420 implements). Video that isn't 8-bit yuv420p or yuv422p, or 10-bit 4:2:0 (below), uses the normal path anyway. Rotated video is converted before it is turned, as OpenCV does, so turned video takes the fast path too. On ARM, where FFmpeg's own conversion rounds differently, it gives x86's results.

10-bit 4:2:0 (yuv420p10le, IYuv420P10Converter, which SwscaleYuv420 implements too) has no unscaled converter in swscale, so x86 scales it: the port follows that path (the bicubic chroma filters, interpolating from where the stream says the chroma sits; the MMXEXT vertical scaler and output; the C output for the last two rows), for even widths and any height, and gives x86's bytes there (checked against FFmpeg 8.1 on x86-64, generated and decoded frames). It exists for other CPUs' results; on x64 it gives the same results as the normal path at about the same speed. It converts a frame's pixels in one pass (IYuv420P10FrameConverter), filtering each chroma row once, as swscale does.

At an odd decoded height swscale can't take its unscaled converter for 8-bit 4:2:0 or 4:2:2 and scales instead, interpolating the chroma from where it sits: the 10-bit port runs those too, on the 8-bit samples widened to 10 bits (swscale's 8-bit and 10-bit horizontal scalers then give the same 15-bit rows), with 4:2:2's one-tap vertical output (yuv2bgr24_1) for 4:2:2 (IYuvOddHeightConverter, which SwscaleYuv420 implements; even widths).

Each converter also says which chroma rows converting a row reads (IYuvChromaRows), so decoding in-process ShotDetector copies, and with Yadif deinterlaces, only the rows it converts, not whole frames.

Copyright (C) 2001-2012 Michael Niedermayer, (C) 2009-2010 Konstantin Shishkov, (C) 2024 Niklas Haas (FFmpeg); yadif (C) 2006-2011 Michael Niedermayer, 2010 James Darnley (FFmpeg); C# port (C) 2026 Jakob Boman.