Change fast rcnn's backbone from resnet to densenet, model cannot converage

Hi, I want to training faster rcnn with densenet as the backbone rather than resnet.
ROI Pooling seems just straightforwardly map roi to feature map. So it should work in many backbone, including densenet. However, when i try to do that, the network cannot converage always. Is it because roi pooling need a specific network to extract features?

Hi @karls

Can’t see a reason why you couldn’t use ROI Pooling with DenseNet. But I can see the selection of feature map having a big difference in the model performance. What feature map were you using from the ResNet backbone? And is there an obvious equivalent in the DenseNet architecture. Or are you just using the last spacial feature map?

With DenseNet the input to a layer is all the other preceding feature maps of a block (concatenated depth-wise), so you might get some benefit of extracting this stacked feature map instead of just looking at single feature map output. Just an idea though.