Community
cancel
Showing results for 
Search instead for 
Did you mean: 
andrea1
Beginner
119 Views

faster-rcnn conversion

Hi all,

I am having some trouble in using faster-rcnn model.

1. Download faster-rnn resenet 101 coco from the zoo model (https://github.com/tensorflow/models/blob/master/research/object_detection/g3doc/detection_model_zoo.md)

2. Convert inference with the following command:

/opt/intel/computer_vision_sdk_2018.3.343/deployment_tools/model_optimizer/mo_tf.py --input_model ../../frozen_inference_graph.pb --tensorflow_use_custom_operations_config /opt/intel/computer_vision_sdk_2018.3.343/deployment_tools/model_optimizer/extensions/front/tf/faster_rcnn_support.json --tensorflow_object_detection_api_pipeline_config ../../pipeline.config --data_type FP32 --reverse_input_channels

3. Run object_detection_sample and got the following error:

[ ERROR ] Can't find output layer named bbox_pred

3.1 If I run the object_detection_ssd_sample it works but the image is resized to 600x600.

Any idea what I am doing wrong?

0 Kudos
3 Replies
Severine_H_Intel
Employee
119 Views

Dear Andrea,

you are doing everything right. faster-rcnn resnet 101 coco needs to be run with the object_detection_ssd_sample. 

Faster RCNN  has been trained in Tensorflow with images resized automatically in the Pre-processor node to 600*600. Therefore, the pre-processing in the inference engine is doing the same, it resizes the images to that size. In case, you want to change this size, please look at the beginning of the pipeline.config and change the min_dimension (see below) and re-run the MO. We do not support non-square inputs yet, we are working on it. 

image_resizer {
      keep_aspect_ratio_resizer {
        min_dimension: 600
        max_dimension: 600
      }
    }
 
Best, 
Severine
andrea1
Beginner
119 Views

Hi Severine,

Thanks for your reply. My resizer params are the following:

  faster_rcnn {
    num_classes: 90
    image_resizer {
      keep_aspect_ratio_resizer {
        min_dimension: 600
        max_dimension: 1024
      }

So I guess because you don't support non-square inputs, the image is resized to the min_dimension? In this case as the "keep_aspect_ratio" is used, if a non-square image is resized to a square image is it padded with zeros?

Thanks

ADua0
Beginner
119 Views

Hello,

I was working on the same model , i got some other related issue to faster Rcnn . I am using object _detection_sample_ssd file to run inference engine,but I am facing some issue. Model is able to detect the object, and give output image but in the output image , no name of the object is coming. But it should come in the output image, like object is car , person etc. I am also attaching the output image I got. I hope anyone can help with this problem.

Reply