DeepLabCut, best practices

Which network to choose?

• Networks available: MobileNets V2, ResNets, EfficientNets
• The difference between them is inference, training speed, and accuracy
• Pretraining -> on ImageNet, but ModelZoo will have animal model-specific training
• MobileNet is faster than ResNet
• ResNet is more accurate than the others. (50 or 101) -> resnet50 is enough for most applications.
• I am using ResNet50.

Which augmentation method to choose?

• Cropping + scaling does a good job, but I used imaug which allows for more diverse augmentation methods.
• So I used imaug option, and it was satisfactory.
• You can also customize the image augmentation.

Training

• Instead of iterating over the network, it is better to train from scratch with all the corrected outlier frames.
• Batch size -> the larger, the better (currently 8)
• Optimizer -> Adam, SGD -> I am using the default one (might be Adam)
• Learning rates are reasonable, no need to change
• Iteration number 400-500k
• Mirror parameter is true for all networks, which helps with generalization.

See more information here and here.

Annotation

• You can copy the previous frame's key point annotations using cmd + c
• To delete a key point, press the scroll wheel in your mouse when the arrow is on the key point.
• Annotate the same place all the time, if the key point is not visible, better to not annotate it
• You cannot change the name of a key point unless you are willing to change all the annotation files you collected, so choose the key point names wisely