Intel® Distribution of OpenVINO™ Toolkit
Community assistance about the Intel® Distribution of OpenVINO™ toolkit, OpenCV, and all aspects of computer vision-related on Intel® platforms.
6257 Discussions

the result for faster_rcnn_resnet_101_coco model is wrong based on OpenVino in MYRIAD Plugin


Hello, everyone 

when  I run the the python sample ( on windows 10 with Intel® Neural Compute Stick 2 powered by the Intel® Movidius™ Myriad™ X , I got following unbelievable and unreasonable result:

the code: res = exec_net.requests[cur_request_id].outputs[out_blob]

the result: [[[ [-1,0,0,0,0,0,0],  [0,0,0,0,0,0,0],  [0,0,0,0,0,0,0].....................

I download the pretrained model from here  

convert the model looks as follows:

C:\Program Files (x86)\IntelSWTools\openvino_2019.1.148\deployment_tools\model_optimizer>python --input_model=frozen_inference_graph.pb --tensorflow_use_custom_operations_config faster_rcnn_support.json --tensorflow_object_detection_api_pipeline_config pipeline.config --reverse_input_channels --data_type FP16

the faster_rcnn_support.json is obtained from  C:\Program Files (x86)\IntelSWTools\openvino_2019.1.148\deployment_tools\model_optimizer\extensions\front\tf

the pipeline.config is along with frozen_inference_graph.pb in the downloaded file.

test model command line  : python -i plane.jpg -m frozen_inference_graph.xml -d MYRIAD

However, the ssd_mobilenet_v1 is successful and result is good. is the the problem of converted  IR model ? or the faster_rcnn_resnet_101 model is currently not  supported ? But I got the message which said this model is supported, you  can see here

Besides, the process for plugin.load(network=net) is too long(at least one minute ) compared to ssd_mobilenet_v1 model.

Additionally,another strange thing is that  the input node in faster_rcnn_resnet_101 model contains two keys(image_tensor and image_info)  while the input node in ssd_mobilenet_v1 model only contains one key(image_tensor ),  is this a problem?


faster_rcnn_resnet_101 model info:
- the input node info 
   - 'image_tensor' # 1 3 300 300 NCHW FP32
   - 'image_info' # 1 3 NC FP32
- the output node info
   - 'detection_output' # 1 1 100 7 NCHW FP32


while the ssd_mobilenet_v1 model info:

- the input node info 
   - 'image_tensor' # 1 3 300 300 NCHW FP32
- the output node info
   - 'detection_output' # 1 1 100 7 NCHW FP32



0 Kudos
1 Reply

Dear Hu, Can,

For faster r-cnn please run the C++ object_detection_demo rather than the SSD one. Please refer to the object_detection_demo document.




0 Kudos