Open Brother-Q opened 3 years ago
I think there is not much thing which can help you to improve the performance. The training set is very small. It's nearly impossible to achieve the same result without loading pretrained model. If you use a more complicate model, it might perform will, but i think it cannot achieve the same level. Moreover, I argue CSRNet. The model's head has big shortcoming.
I tried not to load the pre-training weights of VGG16 and got a MAE more than 200. How do I adjust the parameters in training? Has anyone tried to replace other backbone networks to extract crowd density information?