Vladimir Karadzhov Hello,

these are a few notes I have made I thought you might be interested in the topic.

Okay, Thanks, now I am following you.

Martin Kaloev Thanks, I am working on it, but it is so complex when it gets broken down into smaller tasks, and bringing everything together takes high computing resources, I think starting with only video processing was not possible to get done, the way shown in the repository.

if it happens for you to work on a similar task please let me know.

Okay, that is interesting,

for sure having quality data that covers the task is so important and iterating on the same data will result in overfitting, in machine learning, but I am describing a system trying to make it as close to human as possible so maybe repeating a task could result in a honed and improved skill, and surly to a certain number of times then it needs to get replaced by new data.

and I always clean the data and check for consistency.

thanks for the suggestion.

in case this is in DRL then I am not experienced in that, and currently, I am working on a dataset and this case does exist possibly not sure of the same ratios, but I keep the numbers because the images are for sure not 100% equal in these numbers and that gives training examples for the classification model to be able to improve it's learning, otherwise it would be low-quality data, and always it depends on the application, in some applications it is still could be considered high-quality data such as in time series data like videos when the model does consider the time dimension it is important to keep the data to learn the sequence and transitions which is also considered as information. not like image classification where time does not matter.

but thanks I will keep a note.