Converting a custom ssdlite_mobilenet_v2

Hi all, I'm using the openvino toolkit to convert my custom ssdlite_mobilenet_v2 onnx file to use with my luxonis depthai camera in a modified version of the hellow_world.py example. Here are the steps I've taken:

1) Convert onnx to IR format (.xml and .bin):
python3 mo.py --input_model mb2-ssd-lite.onnx --output_directory --input_shape [1,3,300,300] --data_type FP16 --reverse_input_channels --framework onnx

2) Then convert IR to blob format:
./myriad_compile -m mb2-ssd-lite.xml -o mb2-ssd-lite.blob -ip U8 -VPU_NUMBER_OF_SHAVES 6 -VPU_NUMBER_OF_CMX_SLICES 6

I then modified the hello_world.py example to use my resulting converted mb2-ssd-lite.blob model. Everything goes fine, however, when I go to parse the returned packet it does not have the required format specified here in the openvino documentation.

We expect this:

In [1]: in_nn.getAllLayerNames()
Out[1]: ['detection_out']

But get this instead:

In [1]: in_nn.getAllLayerNames()
Out[1]: ['boxes', 'scores']

Here 'boxes' and 'scores' are the bounding boxes and confidences, but that's all I receive in the packet unlike 'detection_out' which contains [image_id, label, conf, x_min, y_min, x_max, y_max].

** Could it be I doing my conversion process incorrectly in steps 1 and 2 above? **

Sorry for the long post! Thanks inadvance!
Tim

Hi Erik thanks for your input.

I am using the model elsewhere on a different xavier nx environment and it does conform to the output specification given here before conversion with the openvino toolkit:

Output Original Model:

1. Classifier, name: detection_classes. Contains predicted bounding-boxes classes in a range [1, 91]. The model was trained on Microsoft* COCO dataset version with 91 categories of object, 0 class is for background. Mapping to class names provided in <omz_dir>/data/dataset_classes/coco_91cl_bkgr.txt file.
2. Probability, name: detection_scores. Contains probability of detected bounding boxes.
3. Detection box, name: detection_boxes. Contains detection boxes coordinates in format [y_min, x_min, y_max, x_max], where (x_min, y_min) are coordinates of the top left corner, (x_max, y_max) are coordinates of the right bottom corner. Coordinates are rescaled to input image size.
4. Detections number, name: num_detections. Contains the number of predicted detection boxes.

I would like to use this same model on the oak-d. So after the conversion with the toolkit, I expect to get:

Converted Model
The array of summary detection information, name: DetectionOutput, shape: [1, 1, N, 7], where N is the number of detected bounding boxes. For each detection, the description has the format: [image_id, label, conf, x_min, y_min, x_max, y_max], where:

However, instead I'm only receiving in the packet bounding boxes and confidence when using the oak-d api.

Thoughts?