Adjust matching rules videos

[phirestalker]

I have noticed that this software (like all others of its kind) have trouble with videos in low-light scenes. When there is not enough visual information in the scene, the algorithms seem to err towards false positives.

Some videos have an intro that will be matched to all the other videos from that producer.

Is there some way to tell the software to ignore scenes or frames below a certian brightnes threshold? For the other issue, I'm not sure what can be done, other than mark those as non-dulplicates. Is that possibe? Do you have any ideas other than mine how to solve these issues?

PS. This software does most of what I have been looking for. Most software of this kind will only match almost identical videos. yours has found clips of larger videos brilliantly. The only thing that could be better is a way to prefer the larger video and have some kind of saturation map to see if all the scenes are covered by the smaller clips (so they can be deleted).
Thanks for the great program. It has saved me a lot of time and space already.

[scrubbbbs]

Hey, thanks for the feedback.

Some videos have an intro that will be matched to all the other videos from that producer.

use p.vtrim to set the number of frames to ignore from the start/end

Is there some way to tell the software to ignore scenes or frames below a certian brightnes threshold?

This is being done to some extent but I don't have a setting for it.

I like the clip coverage idea, thanks for suggesting.

[phirestalker]

On a related note to the clip coverage. Can this software scan one of those screengrab collection images and match it to a video file? The images with multiple screengrabs every x minutes or whatever?

[scrubbbbs]

On a related note to the clip coverage. Can this software scan one of those screengrab collection images and match it to a video file? The images with multiple screengrabs every x minutes or whatever?

There is -select-grid for this. IIRC it requires some blank space between screengrabs.

cbird -select-grid screengrabs.png -p.alg video -similar-to @ -show

[phirestalker]

I was trying to figure out how to program something like this, and couldn't come up with a good method. Most of the ones I have seen from various tools do not put any space between the frames.
I toyed with using OCR to read the dimensions of the video from the header and detecting the end of the header. Then, dividing the remaining pixels using the found resolution, but that is messy at best. How did you end up implementing it?

[scrubbbbs]

It does need a flat or mostly flat background color between each frame. It finds the borders between thumbnails by looking for horizontal and vertical lines that span the image, which are defined as a series of pixels with close to the same brightness.

I think a similar strategy could be devised where you look for a line using edge-detection methods (but in reverse; i.e. you don't detect edges over the whole image, you trace a line and guess if it is an edge or not).

[phirestalker]

but in reverse; i.e. you don't detect edges over the whole image, you trace a line and guess if it is an edge or not

Yeah, I also wonder if you could start from the outside in and kind of like start with the largest squares and look for the smaller and smaller squares until you find ones where there's multiple squares of the same size.

You said the v trim setting was how many frames to skip in the beginning or end. So does your software use every frame in the video? Will I need to know the frame rate to kind of guess? I know most of the time the frame rate is 30 frames per second.

[scrubbbbs]

Yeah, I also wonder if you could start from the outside in and kind of like start with the largest squares and look for the smaller and smaller squares until you find ones where there's multiple squares of the same size.

I see what you are saying, but finding spanning lines/edges is way easier and should be sufficient as there is the assumption of a regular grid without too much margin/padding around it.

Otherwise you are building a rectangle detector with a constraint on it, something like a rect is only valid if it has enough edges on all sides, has a reasonable aspect ratio, and roughly repeats a whole number of times across/down the image. You could even do some statistics on on all the rects you found to select the most likely solution.

You said the v trim setting was how many frames to skip in the beginning or end. So does your software use every frame in the video? Will I need to know the frame rate to kind of guess? I know most of the time the frame rate is 30 frames per second.

It does decode every frame, but only stores frames that are not too similar to their neighbors. -i.vht sets the threshold for this. So you would need to know the frame rate to guess; most videos are 24-30 fps.

The video indexing is in need of an update to support longer videos and also things like this.

[phirestalker]

I was looking through the help, and there is a directory selection option. To me, it sounds like it is to select what to compare. How can I select the criteria for which image or video to keep. For instance, I want to keep one of a set of exact duplicates if the path has a certain word in it. Also, in general I find it better to keep the photos with more subdirectories, as they are more categorized. Are there any settings for that?

For near-duplicates I would want it to select the one with the highest resolution for a match above say 95 percent. It would be cool if it could move it to the location with the largest path as well.

Sorry, I never want to do anything simple. Thanks for all the help.

[scrubbbbs]

1. cbird -similar compares everything
2. cbird -select-*. -similar-to @ compares selected set to everything
3. cbird -select-* -similar-in @ compares selected set to itself

For instance, I want to keep one of a set of exact duplicates if the path has a certain word in it.

Maybe cbird -dups -without path '~word'

For near-duplicates I would want it to select the one with the highest resolution for a match above say 95 percent

I don't have a percentage scale but you can use a low threshold.

cbird -p.dht 1 -similar -sort-rev res -first

It would be cool if it could move it to the location with the largest path as well.

There is -move but it doesn't support any variable expansion on the destination path. Variable expansion doesn't operate on groups either.

[tehMArs]

Hey, thanks for the feedback.

Some videos have an intro that will be matched to all the other videos from that producer.

use p.vtrim to set the number of frames to ignore from the start/end

p.vtrim that option doesn't work correctly. Even when I use it, (-show) the blank screens on both sides still appear. I've tried various values, ranging from 100 to 3000, to skip the intro, but it doesn't make a difference. I've also noticed that some duplicate removal software has similar options, such as 'ignore white pixels' and 'ignore black pixels', which might be useful in this case.

[phirestalker]

I don't have a percentage scale but you can use a low threshold.

cbird -p.dht 1 -similar -sort-rev res -first

I like this one. The only thing that could make it better is making sure that the file with the largest resolution and file size is the first image. I can't find file size anywhere in prop. I wonder if it is available through ffmpeg metadata?

I have given up on the idea of keeping from a certain folder. The files are too disorganized for this to be useful. My ultimate goal for now is to remove all duplicates that are less quality with -nuke. The only problem is that -p.dht 1 finds images as matches where in one the subject is holding a glass and in the other they are not. In other software where they have a percent match, when it says 100% it can still be resized or recompressed, but it is always the same image. I was hoping -p.dht 0 would do that, but I get no matches with that one. Is there any way to find these simple resizes or recompressions without matching near duplicates?

EDIT: forgot to mention that I am using dct for database size. Do I need a more advanced algo to get what I want, or is it simply not possible yet?

I thought I saw quality score somewhere. What does that take into account? Is it the least compressed and largest resolution file by chance?

[scrubbbbs]

I like this one. The only thing that could make it better is making sure that the file with the largest resolution and file size is the first image. I can't find file size anywhere in prop. I wonder if it is available through ffmpeg metadata?

There is a fileSize property in the current git version. For multisort you can chain properties: -sort-rev res -sort-rev fileSize for example.

Is there any way to find these simple resizes or recompressions without matching near duplicates?

Nope, because they are technically all near duplicates. The only solution to your problem would be a hash code that had more resolution...cbirds is 64 bits but there is no reason it couldn't be wider (well besides the additional overhead).

There is one additional restriction you can add that might remove things you don't want to consider, the template matcher.

cbird -similar -p.tm 1 does a more detailed check for reasonable matches so it could reduce the number you have to go through. The other algos will match cropped and/or rotated images so they're not an option.

There is an attempt at quality scoring in the current version but it is very slow not very good (press "Q" on the UI and it will score the matches). It could be much better, but there is no fool-proof solution besides evaluating each match manually.

[phirestalker]

There is -select-grid for this. IIRC it requires some blank space between screengrabs.

cbird -select-grid screengrabs.png -p.alg video -similar-to @ -show

This gives me the error

@Commands::selectFiles$ select-files: file not found: "@"
@main$ similar-to: empty selection, is "@" a valid path or selector?

[scrubbbbs]

select-grid takes a file list. If the files don't exist it was doing odd things. I just committed a few changes to make it work like select-files. The syntax remains the same.