Training Extensions: how to change the output layer to have confidence score?

[Je6]

Hi, I'm working on OpenVINO training extensions to train object detection model with custom dataset.

I successfully trained model but I found the output layer only includes boxes and labels.

I need confidence score to implement my application.

Before trying to use custom model, I used ssdlite_mobilenet_v2. And this model's output is  DetectionOutput (each detection has the format: [image_id, label, conf, x_min, y_min, x_max, y_max]). 

I wonder why the default structure for object detection(in openvino training extensions) model doesn't have a confidence score.

I've searching related documents and code but haven't found answer..

How can I change this structure to include confidence score?

Thanks.

[Hairul_Intel]

Hi Je6,

Thank you for reaching out to us.

Please provide the following information for further investigation from our side:

• System information (CPU, GPU, OS, Python version, etc.)
• Installation guide that you've followed for the OpenVINO Training Extension.
• Commands used to generate the custom model.

Regards,

Hairul

[Je6]

Hi,

Thank you for responding.

I tried to train models using Semi-Supervised Learning.

• System Information: Dell XPS 15 9520
  • CPU: 12th Gen Intel(R) Core(TM) i7-12700H
  • GPU: GeForce RTX 3050 Ti Mobile, iGPU
  • OS: Ubuntu 22.04.2 LTS 6.5.0-14-generic
  • Python version(in venv): 3.10.12
• I followed "Install OpenVINO™ Training Extensions for users" in this doc. Also tried "Install OpenVINO™ Training Extensions for developers".
• Commands:

$ source .otx/bin/activate
$ otx build Custom_Object_Detection_Gen3_ATSS --train-data-roots ../dataset/person/semisl/coco_format/ --unlabeled-data-roots ../dataset/person/semisl/unlabeled/
$ cd otx-workspace-DETECTION
$ otx train --output outputs/ params --learning_parameters.learning_rate 0.004 
$ otx eval --test-data-roots splitted_dataset/val --load-weights outputs/models/weights.pth --output outputs/
$ otx export --load-weights outputs/models/weights.pth --output outputs/openvino/

So do I need to use the developer version of training extensions? Dose that have more functions? I only see that when I use that one, I can change the code. To resolve the issue I have, should I change the code?

Regards,

[Hairul_Intel]

Hi Je6,

Thank you for providing the information.

We're investigating this issue and will update you on any findings as soon as possible.

Regards,

Hairul

[Hairul_Intel]

Hi Je6,

Thank you for your patience.

Referencing this person-vehicle-bike-detection-2004 model that is based on MobileNetV2 backbone with ATSS head, the "boxes" blob has the following format [x_min, y_min, x_max, y_max, conf].

Here is a code snippet from Object Detection Sample SSD that shows the steps in extracting the confidence score from "boxes" output blob:

In summary, our converted model has already include the "confidence" values inside of the "boxes" blob in the format of [x_min, y_min, x_max, y_max, conf].

Hope this helps.

Regards,

Hairul

[Je6]

Hi,

Yes I realized that there is one more output that indicates labels.

So the structure is:

<bound method PyCapsule.output of <CompiledModel:
inputs[
<ConstOutput: names[image] shape[1,3,736,992] type: f32>
]
outputs[
<ConstOutput: names[boxes] shape[1,200,5] type: f32>,
<ConstOutput: names[labels] shape[1,200] type: i64>
]>>

And in "boxes", there is a confidence score.

Thank you so much for your explanations and response!

Regards,

[Hairul_Intel]

Hi Je6,

Glad to be of help.

This thread will no longer be monitored since this issue has been resolved. If you need any additional information from Intel, please submit a new question.

Regards,

Hairul