Pretrained model Coco with Resnet50 layers

[evinpinar]

Currently, there are two pretrained model weights:

1. Resnet50 Imagenet weights for feature extraction (backbone)
2. MaskRcnn Coco weights for mask-rcnn with resnet101 architecture

Are there any coco trained maskrcnn weights for resnet50 architecture? Or is it okay if I use the resnet101model weights on resnet50 architecture?

[evinpinar]

So far, according to my experiments, it does not work well.

[bpmsilva]

@evinpinar, you could use this tutorial, which uses a native implementation of Mask RCNN in Pytorch

[evinpinar]

Thanks, but I need the weights because i'm modifying this implementation a bit. I guess i will instead move to fb official implementation.