Test the network with the original network ResNet/ MobileNet

Hi @DeepLabCut,

I am working on using DLC technique to train and evaluate the performance of the network with infant data. I am wondering if DLC is possible to test the original network (ResNet/ MobileNet) without adding any infant data at all. That way I can proceed to test my data with the original ResNet/ MobileNet model.

Thank you very much,
Mai

the encoder architecture (ResNet or MobileNet) is only half the model, you need a decoder. Without the decoder, you get no predictions, so no, you need to use training data. If you want to not use a pretrained network, you can train from scratch, see Mathis et al 2019 on Arxiv.

1 Like

Indeed, the (adapted) backbones trained on ImageNet won’t perform pose estimation. Infant’s are also not objects in ImageNet, so they won’t be predicted either.

You could try the full_human network though, which is trained on pose estimation of human adults: http://www.mousemotorlab.org/dlc-modelzoo