Experiments on baseline

[toliyu]

Hi, I'm very impressed by your work. I want to know whether your team has implemented experiments on other baselines, like TopoCount on SANet, CAN, or VGG16 without unet architecture. I've tested some of these networks with TopoCount, but the dice loss remains 0.7-0.8 in the first 30 epochs, and the result is not as good as your original code's. I wonder if you have done tests like this? Maybe I'm doing it wrong? How do you think can I improve this? Looking forward for your reply~

[ShahiraAbousamra]

Thank you for your interest. TopoCount tries to localize each individual in the image. For that, it needs to work on a resolution of the image that allows this. That is why methods that scale down the image to a very small size are not suitable for TopoCount, while the lower resolution is more suited to density-based crowd counting methods. In the paper we use a U-Net with VGG16 encoder. You can try changing the architecture of the encoder in the U-Net.