End to End training with MaskRCNN

[akeaveny]

Hi,

I'm curious if you thought of integrating the ResNet50 backbone and mask branch from MaskRCNN into DenseFusion for end-to-end training? It seems like the ResNet18 network is duplicated for rgb features.

Thanks for sharing your work,

Aidan

[wkentaro]

Yeah, I think it could be more memory efficient.
One problem would be the training experiment would be slower because we jointly train the model for instance segmentation and 6D pose estimation.

[akeaveny]

Thank you for the quick reponse.

I only worked with Matterports Mask R-CNN and Wang's Densefusion before coming across your work. There are a few repos with Mask R-CNN in pytorch which would help with integration. I would have to reconfigure a dataloader for DenseFusion to include GT masks.

One thing I noticed with DenseFusion is that batchsize is limited to one which would slow down training for Mask R-CNN.

Can I also ask if you trained on 640x480 images with a Asus Xtion intrinsics similar to PoseCNN for the YCB dataset?

[wkentaro]

Sorry for the late reply. I missed that.

Can I also ask if you trained on 640x480 images with a Asus Xtion intrinsics similar to PoseCNN for the YCB dataset?

Yes. I use the intrinsic parameters in the YCBVideo dataset.