I’ve been using DLC for 3 or so months now and have tried 2 methods for training and wanted some feedback / had questions.
In the first, I just take any videos I want to label and add them to the training set, and if I have new videos I want to have trained, I just add them to the network and retrain the network later in the project, which updates the network. The analyzed videos from this have been extremely accurate and I have liked what I see. However, it is a lot of work to have to label every video that I want trained.
In the second method, which I initially tried, I put a 10 video training set in and used kmeans, resnet_50 and other default settings to train a network. After that, I loaded different videos than the initial training set into analyze videos to see if the network could be applied broadly, which would save a lot of time from having to train too many videos. However, the analyzed videos of the newly loaded videos were pretty bad (labels were way off and extract outlier frames wasn’t helping).
Mainly, I just want to know which method is ‘normally’ used or if there is another one you guys have found to be better.