I am using a mask_rcnn_inception_v2 model from TensorFlow Object Detection Zoo. I successfully converted it and received .bin and .xml files. When I run mask_rsnn_demo and submit these files to the input, everything works when calculating on the CPU. The model loads quickly and at the output I get an image with masks. However, when I specify a NCS2 as the target device, the model first loads for a long time (this is understandable because the model is large), but after the calculations there are no masks and others results. I tried running this model through python and openvino APIs and at the exit from the network arrays were filled with zeros.
I tried to use both FP32 and FP16 formats.
Example of command which help I made .bin and .xml files:
python /opt/intel/openvino_2020.1.023/deployment_tools/model_optimizer/mo_tf.py --input_shape=[1,800,1365,3] --input=image_tensor --transformations_config=/opt/intel/openvino_2020.1.023/deployment_tools/model_optimizer/extensions/front/tf/mask_rcnn_support.json --tensorflow_object_detection_api_pipeline_config=/home/alexey/PycharmProjects/teaching/segment_plate/freeze/pipeline.config --input_model=/home/alexey/PycharmProjects/teaching/segment_plate/freeze/frozen_inference_graph.pb --data_type FP32 --progress
Thanks for reaching out.
Could you please do the following?
- Try using these flags: --data_type FP16 --reverse_input_channels. And tell us if the issue persists.
- Provide us your files for us to test on our end.
Is there any update on this post. I am Also working on Intel NCS 2.
I am Running SSD Mobile Net model successfully with NCS2 and Raspberry Pi.
I need Mask RCNN to run in same setup. Any Idea in possibilities.