flyinglynx
commented
2 years ago Hi. In this work we use ResNet-50 for fair comparisons with previous works CFOCNet and FamNet.
I have not tested other deeper backbones. Sorry that I really have no idea how backbones like swin and convnext will perform. I tested ResNet-18 and the MAE on validation set increases to around 18. I notice a recent work (link) obtains better performance but using only ResNet-18, unfortunately the code is not released.
Whether better backbone can improve performances on counting tasks is doubtful. According to my experimence on class-specific counting tasks, simple VGG backbone may performs better than ResNet. Maybe you can test other backbones and see the impact of backbones on class-agnostic counting.
I conducted the test and your work has achieved good results. I have some questions to ask. As described in the article The feature backbone consists of the first 4 blocks of ResNet-50 [12], which outputs the feature maps of 1024 channels. Why use the ResNet-50 backbone instead of another deeper network? Is there any specific reason to use ResNet-50? Thank you!