example video switches individuals

[MCMaurer]

Hi there!

I've been playing around with tracktor and I'm really excited to apply it to some videos! I've been looking at some of the examples and I noticed that in the tracked zebrafish video, there are some instances when the individuals get switched. I noticed it at frames ~1300 and ~1500 in zebrafish_video_tracked.mp4. Both switches occur between individuals that are quite far apart, not just when two individuals pass close to each other.

I was wondering how often you run into this problem and how best to go about solving it.

Thanks!

[vivekhsridhar]

Hello,

You are right that the identities switch sometimes between further away individuals. This is a result of the hungarian algorithm implemented for identity maintenance. Since the algorithm tries to minimise the overall distances between successive frames, it sometimes swaps around IDs of multiple individuals such that this costs function is minimised.

A software like IDTracker (with a much more sophisticated identity maintenance algorithm) is suitable for this. The advantage of tracktor however is that it has a much better detection rate. The best tracking performance will hence incorporate a combination of these two software.

At some point, I will try and release this as an addition to Tracktor but this might be a while.

Hope this helps.

Vivek

[MCMaurer]

Hi Vivek,

Thank you so much for the response! I'm really excited to use Tracktor on some of my videos with single individuals, and I figure I'll give it a shot with groups too!

One other thing- I noticed that in the CSV file, right around the frames where the switches occurred, the pos_x and pos_y values went from having many decimal places to having only integer values. Is this a possible indicator that Tracktor is struggling to maintain individual IDs?

Here's what I'm thinking- if there is some indicator in the CSV that a switch may have occurred, such as a change in the number of decimals, I should be able to easily pick out those spots. I could then generate a list of frames with potential switches, and then go look at those points in the video to see if a switch happened. If a switch did occur, it should be simple enough to modify the CSV to correct for the identification mismatch, saying "alright, any individual flagged as A after this point is actually individual B". This would be pretty awesome!

Again, thanks for all the work on Tracktor, I think it's amazing and I'm really excited to use it more!!

[vivekhsridhar]

Hey, I don’t see why the number of decimals should change when identity switches occur. There’s nothing in the code that should cause this. But that said, if this is a consistent pattern that you notice, then that sounds like something that could work. Although I’m skeptical that that would be a consistent pattern. What you could also do is to get individual speeds. Identity jumps will be accompanied by jumps in speed (since its calculated as the distance between successive frames). You could use some threshold speed which your fish wouldn’t exceed to identify these jumps. Hope this helps Vivek ------------------------------------ Vivek Hari Sridhar website: www.vhsridhar.wordpress.com <http://www.vhsridhar.wordpress.com/>

…

On 07-Feb-2019, at 05:29, MCMaurer ***@***.***> wrote: Hi Vivek, Thank you so much for the response! I'm really excited to use Tracktor on some of my videos with single individuals, and I figure I'll give it a shot with groups too! One other thing- I noticed that in the CSV file, right around the frames where the switches occurred, the pos_x and pos_y values went from having many decimal places to having only integer values. Is this a possible indicator that Tracktor is struggling to maintain individual IDs? Here's what I'm thinking- if there is some indicator in the CSV that a switch may have occurred, such as a change in the number of decimals, I should be able to easily pick out those spots. I could then generate a list of frames with potential switches, and then go look at those points in the video to see if a switch happened. If a switch did occur, it should be simple enough to modify the CSV to correct for the identification mismatch, saying "alright, any individual flagged as A after this point is actually individual B". This would be pretty awesome! Again, thanks for all the work on Tracktor, I think it's amazing and I'm really excited to use it more!! — You are receiving this because you commented. Reply to this email directly, view it on GitHub <#8 (comment)>, or mute the thread <https://github.com/notifications/unsubscribe-auth/AM_9jUvKGKevigTObLPfe0TKyp7KIcQcks5vK6uwgaJpZM4afUot>.