Custom object classification with Yolo and Deepstream

Accelerated Computing Intelligent Video Analytics DeepStream SDK
1

I learnt that YOLO with deepstream has way better performance than YOLO with OpenCV. How can I train a custom dataset? Would be great to get a stet by step guide as I am a newbie. Thanks a lot

3

Hi,

You can use YOLO author’s API to retrain the model.
For more detail, please check this page:

Training YOLO on VOC

Thanks.

4

Thanks. But how do I connect it to deepstream? What should I modify?

5

Hi,

After training the new YOLO model with darknet, please follow our document to customize the parameters:

Then you can run the model with our sample directly:

/opt/nvidia/deepstream/deepstream-4.0/sources/objectDetector_Yolo/

Thanks.

6

Hi, thanks for the reply. I am confused with step 3. Where should I get these parameters and what exactly should I change.

Without step 3 modifications I managed to run deepstream app, however, it does not draw bounding boxes and does not classify. Just outputs the input video. I assumer, I need to modify step 3, could you please tell me where should i get modifications? Thanks a lot!

7

If you retrain a model with specific ground truth anchors and clusters, you have to modify them in the following function, mentioned in the step 3.

If you change only the training dataset, but not the configuration, all you need is to replace the weights file to your own.

8

Hi,

Please check the document shared above.
It will tell you how to update the configure based on your cfg file.

Thanks.