split-video sometimes cut at wrong frames/length

[Werve]

Description:

With one particular video file (which I think I can share privately if needed) I noticed that created cut scene files generated by this command did not have the correct length in frames reported in the SceneTimings.csv file. Yet with recoding, in this case lossless, according to the doc (https://trac.ffmpeg.org/wiki/Seeking) the ffmpeg -ss commands should ensure accuracy.

scenedetect -i “%_InputFile%” -b pyav load-scenes -i SceneTimings.csv -c “Start Frame” split-video -a “-map 0:v -c:v libx264 -preset veryfast -crf 0 ” -f $SCENE_NUMBER.mkv

(With SceneTimings.csv generated by scenedetect -i “%_InputFile%” -b pyav -s stats.csv --min-scene-len 10s detect-adaptive --weights 1 1 1 list-scenes -s -f scenesTimings.csv but it should not affect the behavior)

I also tried split-video with -m flag and manually running the ffmpeg commands with the values reported in the SceneTimings.csv file with different seek methods which I report in the following table where there is the number of frames per scene calculated with the command:

ffprobe -i “%SceneNumber%.mkv” -select_streams v:0 -count_frames -show_entries stream^=nb_read_frames -v quiet -of csv^=“p=0” >> framelist.txt

Scene n.	SceneTimings Frames	scenedetect -m	Diff	scenedetect mkv out	Diff2	scenedetect mp4 out	Diff3	ffmpeg mkv	Diff4	vf select	Diff5
1	683	684	1	679	-4	679	-4	653	-30	655	-28
2	358	380	22	359	1	359	1	257	-101	257	-101
3	306	304	-2	306	0	306	0	287	-19	289	-17
4	334	309	-25	333	-1	333	-1	316	-18	314	-20
5	286	355	69	289	3	289	3	281	-5	284	-2
6	573	506	-67	572	-1	572	-1	255	-318	254	-319
7	277	330	53	275	-2	275	-2	244	-33	243	-34
8	314	284	-30	316	2	316	2	228	-86	230	-84
9	324	301	-23	324	0	324	0	254	-70	254	-70
10	269	269	0	269	0	269	0	251	-18	251	-18
11	301	342	41	299	-2	299	-2	257	-44	257	-44
12	328	288	-40	331	3	331	3	344	16	346	18
13	305	351	46	302	-3	302	-3	274	-31	274	-31
14	268	219	-49	268	0	268	0	333	65	334	66
15	257	262	5	259	2	259	2	291	34	289	32
16	248	304	56	246	-2	246	-2	264	16	266	18
17	266	205	-61	266	0	266	0	256	-10	257	-9
18	268	307	39	268	0	268	0	270	2	271	3
19	271	304	33	271	0	271	0	284	13	285	14
20	286	214	-72	287	1	287	1	247	-39	247	-39
21	352	352	0	352	0	352	0	247	-105	247	-105
22	305	346	41	307	2	307	2	441	136	440	135
23	274	284	10	272	-2	272	-2	294	20	294	20
24	270	219	-51	269	-1	269	-1	243	-27	245	-25
25	295	313	18	296	1	296	1	398	103	398	103
26	302	284	-18	301	-1	301	-1	360	58	358	56
27	265	324	59	265	0	265	0	354	89	355	90
28	261	228	-33	261	0	261	0	262	1	263	2
29	460	436	-24	461	1	461	1	335	-125	335	-125
30	260	258	-2	258	-2	258	-2	597	337	599	339
31	414	414	0	414	0	414	0	312	-102	311	-103
32	375	377	2	377	2	377	2	370	-5	371	-4
33	379	451	72	378	-1	378	-1	431	52	431	52
34	273	228	-45	272	-1	272	-1	314	41	315	42
35	354	380	26	355	1	355	1	447	93	447	93
36	624	588	-36	622	-2	622	-2	1405	781	1400	776
37	332	380	48	333	1	333	1	266	-66	264	-68
38	387	380	-7	386	-1	386	-1	546	159	547	160
39	450	456	6	450	0	450	0	260	-190	262	-188
40	467	401	-66	467	0	467	0	294	-173	292	-175
41	1486	1486	0	1486	0	1486	0	305	-1181	305	-1181
Sum	15107	15103	-4	15101	-6	15101	-6	14327	-780	14336	-771

I noticed that the split made by scenedetect split-video using ffmpeg does not have the total frames as that of the original video, and neither do the manual ones using ffmpeg (-ss before or after -i, and -vf select="between(n,start_frame_num,end_frame_num) ).
The split that comes closest is with mkvmerge but it does not precisely cut each scene in the chosen timestamps.

Environment:

scenedetect version

[PySceneDetect] PySceneDetect 0.6.5
[PySceneDetect] Loading user config file:
  C:\Users\A\AppData\Local\PySceneDetect\scenedetect.cfg

System Info
------------------------------------------------------------
OS               Windows-10-10.0.19045-SP0
Python           CPython 3.13.0rc1
Architecture     64bit + WindowsPE

Packages
------------------------------------------------------------
av               13.1.0
click            8.1.7
cv2              4.10.0
imageio          2.36.0
imageio_ffmpeg   0.5.1
moviepy          2.1.1
numpy            2.1.3
platformdirs     4.3.6
scenedetect      0.6.5
tqdm             4.67.1

Tools
------------------------------------------------------------
ffmpeg           7.1-full_build-www.gyan.dev
mkvmerge         v88.0 ('All I Know') 64-bit

Media/Files:

The SceneTimings.csv :

scenesTimings.csv

If it is needed, perhaps to replicate or create a test, I can privately send the video file somehow.

EDIT: After looking at the mediainfo of the input file, it reports variable frame rate for the video.
Could it explain the mismatch between the frames written in SceneTiming.csv and the actual frames in the cuts?

[Breakthrough]

EDIT: After looking at the mediainfo of the input file, it reports variable frame rate for the video.
Could it explain the mismatch between the frames written in SceneTiming.csv and the actual frames in the cuts?

Yes that would explain it! VFR videos aren't well supported right now, this is being tracked in #168. I hope to eventually overhal the timecode APIs so that VFR works better with PySceneDetect.

For now, the best workaround is, unfortunately, to re-encode the input to be CFR.