High false positives in partial clip detection when matching short clips to merged full video

[GQH666666]

Hi Maintainers,

I'm using Video Duplicate Finder and encountering issues with partial clip detection in a specific scenario:

1. I have one high-quality video that is merged from two separate source videos.
2. I also have several short video clips, each of which is a segment cut from a part of the merged full video.

My question is:
Can Video Duplicate Finder reliably recognize these short clips and match them to their corresponding positions in the merged long video?

Currently, when I enable partial clip detection mode, the false positive rate is extremely high, making the results completely unusable for this use case.

Thank you for your great work on this tool!

[GQH666666]

[0x90d]

Hi @GQH666666,

Possibly addressed - a few partial-clip improvements landed shortly after you posted, the most relevant being commit 0be846f, which adds a visual gate: after an audio match, VDF now decodes 1–3 frames at the matched offset on both the clip and the source and drops the pair if they don't also match visually. This was specifically aimed at the "shared-audio-track but unrelated visuals" false-positive class, which sounds close to (though maybe not identical to) what you're describing.

Two other fixes from the same window may also help your case:

• 1d738aa - drops singleton groups where one short clip spuriously matches multiple unrelated sources.
• 252aa1c - skips silent audio tracks (a common false-positive source).

Could you grab a recent build (or the next 3.0.x release) and re-run on the same dataset? Relevant settings under Partial clip detection:

• Require visual match - leave on (default).
• Visual match threshold - default 0.85. Try 0.90 if you still see false positives, or 0.80 if you're losing real matches.
• Min similarity - default 0.80. Raise it (e.g. 0.85–0.90) if your clips' audio is generic/music-heavy.
• Enabling pHash generally tightens the visual gate too.

If false positives are still excessive after that, please reply with:

1. A rough idea of clip duration vs. merged-video duration,
2. Whether the audio across clips is similar (same soundtrack/dialogue),
3. The exact settings you used,

and I'll take another look.

[GQH666666]

My main use case is to match short video clips and find which full-length original video they come from. Could you please advise me on the best parameter settings for this scenario?

[0x90d]

Thanks for the screenshots. For "short clip → which long merged source did it come from", the default settings are actually working against you, because Min clip ratio assumes the clip is a sizeable fraction of the source. Try this profile:

Settings to change (under Partial clip detection):

Setting	Default	Recommended for your case
Min clip ratio (%)	10	2–5
Min audio similarity (%)	80	88–90
Require visual match	on	leave on
Min visual similarity (%)	85	90
(Main settings) Use pHash	—	enable

[GQH666666]

The feature that lets me trace short clips back to their original long videos is fantastic! Thank you so much for your work and quick replies. Everything is working perfectly now.