Along with my team, we are doing project 2 in a neuroscience lab at EFPL. Our project consists in running an analysis on several videos of patients performing a robotic task thanks to a deep learning image recognition software made in EPFL (DeepLabCut). My question is not specific to our project, but applies to many different machine learning applications, and it deals with the video analysis. It seems that the videos are captured from many different angles, each time changing the way our coordinates will be extracted from the footage and therefore changing the consistency of our dataset. The goal of the project is to try and diagnose patients suffering from Parkinson's only by analyzing the movement of the hand.
Here is the thing : would it be possible despite having an inconsistent dataset, to conduct a satisfying analysis of the data ? Would it mean building many different models ? How would we be able to train them while none of the videos are relying on consistent settings ?
This is quite concerning as we are not yet sure we will be able to have any result at all, since we still are working on the preprocessing step, and wanted to have somewhat of an idea as to the feasibility of the project.
I've been assigned to the project "Supervised classification of fly behaviors from pose tracking data" which may be a bit related to what you describe. I think it's better to take it to PM -- could you please message me in Discord and describe in more detail the dataset you have (what are the inputs / outputs) and the way how you are going to use DeepLabCut?
Is our project feasible ?
Hello everyone,
Along with my team, we are doing project 2 in a neuroscience lab at EFPL. Our project consists in running an analysis on several videos of patients performing a robotic task thanks to a deep learning image recognition software made in EPFL (DeepLabCut). My question is not specific to our project, but applies to many different machine learning applications, and it deals with the video analysis. It seems that the videos are captured from many different angles, each time changing the way our coordinates will be extracted from the footage and therefore changing the consistency of our dataset. The goal of the project is to try and diagnose patients suffering from Parkinson's only by analyzing the movement of the hand.
Here is the thing : would it be possible despite having an inconsistent dataset, to conduct a satisfying analysis of the data ? Would it mean building many different models ? How would we be able to train them while none of the videos are relying on consistent settings ?
This is quite concerning as we are not yet sure we will be able to have any result at all, since we still are working on the preprocessing step, and wanted to have somewhat of an idea as to the feasibility of the project.
Thank you
Camil
Hi Camil,
I'm not sure about the exact name of your project, so I'm not sure who is your assigned TA (full list: https://docs.google.com/spreadsheets/d/1tXqsS9jakdJk-FOlQMAhU3iQ0Iq_xidahFxC__sazdQ/edit#gid=2101315809).
I've been assigned to the project "Supervised classification of fly behaviors from pose tracking data" which may be a bit related to what you describe. I think it's better to take it to PM -- could you please message me in Discord and describe in more detail the dataset you have (what are the inputs / outputs) and the way how you are going to use DeepLabCut?
Thanks,
Maksym
Add comment