Maskrcnn (segmentation) using a converted model in c++ without DeepStream

Hi,

Using maskrcnn from tlt v2, on jetson, following the notebook training in the user guide.
— NOT USING DEEPSTREAM —

  1. Say I managed to convert (using tlt-converter) the model . How do I parse the results in my c/c++ code ? That is, I enqueue the input to the model, and copy the results (say using cudaMemcpyAsync) to host buffer. What format is the data? (in classification its easy - prob/class, in detection also ok - class,prob and bb) Is there any documentation on this ? Just to make clear - I am not using DeepStream.

  2. Another quick q - can I change the input size of the model in the maskrcnn_train_resnet50.txt ? I mean, is it size insensitive like YOLO (up to the x32 multiplication ) ?

Thanks for the help !

  1. See https://docs.nvidia.com/metropolis/TLT/tlt-getting-started-guide/text/deploying_to_deepstream.html#generating-an-engine-using-tlt-converter

For classification use: predictions/Softmax.

  • For DetectNet_v2: output_bbox/BiasAdd,output_cov/Sigmoid
  • For FasterRCNN: dense_class_td/Softmax,dense_regress_td/BiasAdd, proposal
  • For SSD, DSSD, RetinaNet: NMS
  • For YOLOv3: BatchedNMS
  • For MaskRCNN: generate_detections, mask_head/mask_fcn_logits/BiasAdd
  1. Yes, you can change the input size, but currently, the width or height should be multiples of 64.
    Refer to Error running MaskRCNN inference after custom training - #6 by hyperlight

Great, thanks.