Crating a Deeplabcut project with existing images and labels from


I have a set of frames and a .csv with bodypart coordinates (per frame) which were labeled outside of DeepLabCut.
I do not have the videos from which these frames are from.
How should I create a DeepLabCut project folder, with everything needed for training, in this scenario?


Hi, Dan:
That depends on what the labels were created in, but it is possible. Where are the labels from?

However, you say you don’t have the videos. Do you have still images? Without videos or images, there is nothing to train the neural network on.

Thanks Brandon,

The labels were created in DeepPoseKit and I have a pandas dataframe with these.

Certainly the convnet needs images to train; I have the images of only the labeled frames, and not the whole videos from which these frames were extracted. I.e., I don’t have video paths to provide to config.

Thanks again,


Hi @Dan_Biderman, this post may be helpful. It is not ideal, but this is a simple hack to make DLC believe the original video is present. It is important the ‘fake video’ path appears in the config.yaml so that the labeled-data folder is correctly found. The last thing to verify will be that your .csv file does obey DLC data format and to convert it to a h5 file. Hope this helps! :slight_smile:

thanks, I’ll report back soon!

Hi Dan, this is covered fully on the DLC wiki: Using labeled data in DeepLabCut that was annotated elsewhere · DeepLabCut/DeepLabCut Wiki · GitHub

Hope that helps!