Closed imMid-Star closed 10 months ago
In MaskRCNN, we use the original channels_list. In order to align with Resnet50, we added 1x1Conv at the beginning of the Neck to reduce the number of input channels by 2x to fit the original channels_list. This is just a rough alignment and you can explore a more efficient structure for ResNet.
Hello, according to Table 4 in the paper, you used GD instead of the original FPN after ResNet50 in MaskRCNN. Can you show the config file of this experiment, especially the channel dimensions? I think ResNet50 is a more common backbone than others and more people may want it.