Deeplabcut General Questions - Repost since idk if anyone saw this last time

Hi guys, I’ve been learning a lot about DLC and have just a couple varying, general questions that I was hoping someone could answer.

  1. What is the function of plot trajectories? When and why should I be using it?
  2. I have been trying to create a successful network for a worm for quite some time now but have been a bit unsuccessful. Is it even possible to produce a successful network for a worm (since other than the head and the tail, most of the body is indistinguishable)?
  3. From the first minute I begin working on making a network to it being finished, is it reasonable to expect a successful network within 24 hours without having extracted outlier frames, or is it the norm to have to extract outlier frames many times before having a successful network?
  4. I’m not sure if this has been asked before, but how do I send a config file to another computer with a different path?
  5. Lastly, is it possible to take the x and y points that we get from the analyzed videos csv file and use those to replot the labels in another program? If so, how and which program?

Thanks in advance and have a great day!

Hey, I’m still fairly new to DLC, but I can take a stab at some of your questions.

  1. The general rule is that if you can visually identify what you want to track, DLC should work. I’m a little unclear what you mean by the body of the worm being indistinguishable (perhaps you could post a representative image).

  2. This probably depends on the properties of your specific images and the type/number of training images you provide, but in my experience working with videos of dark mice on a bright background, I haven’t needed to extract outlier frames - I was happy with my results.

  3. You should be able to move your project folder to a new location - all you have to do is change the “project path” parameter in your config.yaml file. I haven’t tried moving to a different computer entirely, but I think this should still work.

  4. Yes, a csv file can be read by many programs, including excel and matlab (check out the csvread function).

Hello! I found a lot of your answers really helpful, most of this makes sense. For question 2, I don’t have an image on me but you can imagine with worms, the only bodyparts distinguishable are the head and the tail since they are different shapes. The actual body is all the same, so when labeling the bodyparts, I see in analyzed videos that the bodyparts are usually pretty inaccurate. I think in one of my last posts, I posted a video showing this along with the worm if you’d like to check that out.

Ah I see - I went and saw that video, but unfortunately, I don’t think I have any insights to offer beyond what the other commenters posted. Multi-animal tracking is something I’m still trying to figure out, but I haven’t seen your specific issue before. Maybe change the lighting/recording parameters to increase the contrast of your videos? In any case, best of luck!

1 Like

Hey @Quick_Science! I will try to complement @uday’s great answers.

  1. Plot trajectories is very useful as a diagnosis tool to qualitatively evaluate detection quality. Have a look here, this is explained very well. I would add that the xy plot at the top left makes transient jumps in detections very flagrant (you’ll see colored dots way off their clusters).

  2. I cannot agree more with the rule of thumb above. We have seen backs of mice labeled with ~5 keypoints along the spine to the base of the tail and it does work very well. But in your case it may be advantageous to use the multi-animal version (even on a single worm), its the part affinity fields (which, simply put, encodes the position and orientation of ‘limbs’ (or links between pairs of keypoints).

  3. I do not believe the norm is to extract outlier frames many times, although I think that a good practice is to do it once. This should be sufficient for the model to generalize better after retraining. Using a GPU (or Google Colab), you could definitely get something good looking in 5-6 hours. But the best here is really for you to try; there are notebooks available that you could simply run cell by cell :slight_smile:

Your answer is really helpful @jeylau , thank you!

I’ve been using the GUI for my project thus far,because I’m not too experienced with colab and jupyter. Would I still be able to use the multi-animal version doing this? If so, how? I just went on the DLC GitHub and installed the latest version of deeplabcut, but I guess that wasn’t correct?

Absolutely! You get to select whether you want a single- or multi-animal project right when opening the GUI. Cloning the repo is fine, no worries :slight_smile:

1 Like

Thanks! I’ll try with multi-animal project and let you know if that doesn’t work out. Thanks again for taking the time to help out!