Video similarity check: which threshold to set
The similarity check compares a finished copy with the original. It also compares the copy with the other copies of the same video. You set the similarity threshold yourself. The app deletes or recreates a copy above the threshold. You choose the action yourself too. By default the check is off. It is turned on on the "Export" page.
What exactly the similarity check compares
The tour inside the app describes it like this:
A VUB speciality: the built-in similarity check. It compares the finished videos with the original and with each other. Copies that are too similar can be created again or deleted. The check is off by default.
The tooltip of the setting adds three points. They are the same for all check methods:
- "All methods take into account the similarity to the original".
- "All methods take into account the highest similarity to the other copies".
- "If the similarity is above the threshold with at least one earlier copy of the same video, the action from the item below is applied".
The app compares only copies of one source video with each other. The check works during batch video creation. This is creation with the "Start" button. A preview does not go through the check.
How to turn on the similarity check
- Open a project. The "OUTPUT" section on the rail on the left, the "Export" page.
- In the "Similarity check" field choose one of the methods: "AI check", "Video indexing v1" or "Video indexing v2". The value "No" turns the check off.
- Set a value in the "Similarity threshold" field. The threshold is set in percent, from 1 to 100.
- Choose a value in the "If similarity is above the threshold then" field.
- The "Delete and recreate" action adds the "Video re-creation attempts" field. Enter a number in it.
While "No" is selected in the "Similarity check" field, the "Similarity threshold" and "If similarity is above the threshold then" fields are inactive. The "Video re-creation attempts" field is visible only with the "Delete and recreate" action.
What the similarity threshold is and which way is stricter
The app applies the action when the similarity exceeds the threshold. Hence the rule: the lower the threshold, the stricter the check. The lower the threshold, the more copies exceed it. The tour says it directly: "The lower the similarity threshold, the stricter the check".
The field tooltip gives two reference points of the scale:
| Value | What it means |
|---|---|
| > 99% | the videos are identical |
| < 10% | two different videos |
It also says: "the threshold is chosen individually for each video hosting or social network". And also: "with too low a threshold it may delete every created video".
A threshold of 1 together with deleting the copy is the extreme case. The "Check before start" window then shows the warning "The similarity threshold is at its lowest". Almost any finished copy falls under such a threshold.
In a new project the field is set to 90%. The field itself is inactive, though, because the check is off. Choose a check method, and the field becomes available.
What happens to a copy above the threshold
You choose the action in the "If similarity is above the threshold then" field. There are three options.
| Action | What happens to the copy |
|---|---|
| "Delete and recreate" | The copy is deleted. The app creates a new one in its place with new random parameters. |
| "Delete" | The copy is deleted. No new one is created in its place. |
| "Do nothing" | The copy stays in the save folder. It has no separate mark in the table. The row ends with the "Done" status. Such a copy can be recognized only by the numbers in the "Similarity to original" and "Similarity to copies" columns: they are above the threshold. |
You set the number of attempts in the "Video re-creation attempts" field. The default value is 10. The tooltip describes the end of the attempts like this: "The attempts end when the limit is used up or the created video fits the similarity threshold". The attempts may run out. The app then deletes the copy above the threshold.
While the app recreates a copy, the "Status" column shows "Recreating the video (attempt 1/10)". When videos are created in the cloud, it shows "Recreating in the cloud (attempt 1/10)". Both similarity columns show "Recreating..." during this time.
How the check methods differ
| Method | What it takes into account |
|---|---|
| "No" | No check is done. |
| "AI check" | Artificial intelligence trained on human ratings. It computes the peak signal-to-noise ratio PSNR. It takes into account the structural similarity index SSIM, the detail loss index, object motion over time, judder and flicker. |
| "Video indexing v1" | Dominant colors in the frame, brightness gradients and contrast. The overall color range of the video and how it changes as the video goes on. |
| "Video indexing v2" | Everything v1 takes into account, plus object geometry, human facial features, silhouettes and outlines. It takes the frame composition into account: tells the background from the foreground. Resistant to simple color manipulations. |
The v2 method is built on top of v1. They compute the color part the same way. The v2 method adds the analysis of frame geometry and composition.
The tooltip also answers the question "which one to choose": "Find the method you need by testing it for your own case". It also has a caveat: "Different methods can give different scores". It is not worth carrying a familiar threshold over from one method to another.
In versions before 1.9.4 these settings had other names. The "AI check" method replaced the "Check similarity to the original and the other copies" setting. The "Video indexing v1" method took the place of the "Index videos" setting. The app moves old projects to the new values when they are opened.
What the results table shows
The report table has two columns: "Similarity to original" and "Similarity to copies". While the check is off, both show the "-" sign.
| Value in the column | What it means |
|---|---|
| "Waiting for video..." | The copy is still being created; there is nothing to compare yet. |
| "Comparing with the original" | The app compares the copy with the source video. |
| "Comparing with copy 1" | The app compares the copy with one of the earlier copies. Only "AI check" shows this label. The "Video indexing v1" and "v2" methods compare copies with each other right away and do not show a separate step. |
| "First copy" | There is nothing to compare with: there are no other copies of this video. The app does not count copies above the threshold in the comparison, even if they stayed in the folder. |
<- Above the threshold | The copy turned out to be above the threshold already in the comparison with the original. The app did not compare it with other copies. The arrow points to the neighboring column with the number itself. |
| A number with a percent sign | The finished similarity score. |
| "Error", "N/A" | The score could not be obtained. |
What it costs in time and resources
The check runs on your device. This is also true for the "Cloud" mode. The server creates the video. The app downloads it to the save folder. Your device compares the files.
All three methods read the whole video. So they add time to each copy. They also load the processor. After that the methods differ. "Video indexing v1" and "v2" read each file once. They save its fingerprint. So comparing with finished copies costs almost nothing. "AI check" compares video with video. Each comparison is another pass over the file. As a video gets more copies, checking each next copy takes longer.
The "Delete and recreate" action is a separate matter. Each attempt is one more video creation. When videos are created in the cloud, each attempt charges credits again. So the "Check before start" window shows the line "With recreations (up to 10 attempts)". It reserves the amount for the worst case.
Frequently asked questions
Why are the "Similarity threshold" and "If similarity is above the threshold then" fields inactive?
Because "No" is selected in the "Similarity check" field. Please choose one of the three check methods. Both fields become active. The "Video re-creation attempts" field appears with the "Delete and recreate" action.
The check deletes all created copies. What do I do?
Most likely the threshold is too low. The app applies the action to all copies above the threshold. Please raise the value in the "Similarity threshold" field. The "Check before start" window gives the same advice in the warning "The similarity threshold is at its lowest".
Does the similarity check work when creating videos in the cloud?
Yes. The video creation method does not affect the check. The comparison rules are the same on the device and in the cloud. The comparison itself runs on your device. More about the modes in the article Video creation on the device and in the cloud.
What to choose: "AI check" or "Video indexing v2"?
There is no universal answer. The tooltip in the app answers the same way. The method is chosen by testing for your own case. Different methods give different scores for the same video. The threshold is chosen anew for each method.
Is the preview checked?
No. The check works only during batch video creation. This is creation with the "Start" button.