MaskRCNN ONNX export issue

Hello community.

I have exported MaskRCNN to ONNX, but my problem is that, I need convert this model into tensorrt.

The problem is:

  1. TensorRT does not support dynamic inputs and outputs;
  2. MaskRCNN have dynamic inputs and outputs.

I can fix MaskRCNN inputs to a certain size, but once that done, the outputs were also fitted.

Does anyone knows:

How to get a fixed output and a fixed input?? (Every images has different objects, so the final output is not exactly same).

I want a fixed output dimension, such as RPN output fixed in 6000 candidates. So that we can do the resting by ourseles.

1 Like

I checked this problem recently and I found your questions and answers here and there. I seems that we met the same problem for fasterrcnn and maskrcnn. It seems a deadend so far.