Tensor RT4 for Mask RCNN

Has anyone successfully implemented Mask RCNN on Tensor RT to boost inference speeds? If yes, how much an improvement was it and which Mask RCNN model did you implement to do so? I’m trying to convert this [model][http://download.tensorflow.org/models/object_detection/mask_rcnn_resnet50_atrous_coco_2018_01_28.tar.gz] to Uff but Tensor RT does not support the mask layers.

If you check their developer guide https://docs.nvidia.com/deeplearning/sdk/pdf/TensorRT-Developer-Guide.pdf You see FastRCNN has been implemented, and it does give you the bounding boxes, I think you’d have to implement the RoiAligh plugin layer to deploy with TensorRT.

I’m currently trying to do so, so try to check out their dev guide and the FastRCNN example, it’s a decent starting point.