Semi Supervised Learning using Jetson Nano

[swarnika.sarangi]

In my project, I need to detect food label on cardboard box using the MIPI camera on Jetson Nano. I followed the tutorial for Object detection with our own dataset:

dusty-nv/jetson-inference/blob/master/docs/pytorch-collect-detection.md

I was able to detect the food label successfully on the cardboard box.
Now, I need to use Semi Supervised learning in such a way that when an unlabelled dataset is provided to the model it should use the pre-trained ssd model to detect the food label. Is there any way to do this?

[dusty_nv]

@swarnika.sarangi

Now, I need to use Semi Supervised learning in such a way that when an unlabelled dataset is provided to the model it should use the pre-trained ssd model to detect the food label. Is there any way to do this?

Hi @swarnika.sarangi, do you mean that your SSD model will need to detect new object classes that it wasn’t trained on before? To do that you would need to add new annotated images of these additional classes to your dataset and retrain your model.

Or do you mean that you want to detect the objects you already trained on, but with backgrounds different than the cardboard box? You can experiment with your model during inference to see how robust it is against different backgrounds and lighting conditions. I would recommend collecting more data with different backgrounds and re-training your model. The train_ssd.py script isn’t set up for semi-supervised learning I don’t think.