Help understanding EfficientDet-lite output classes

TensorFlow
,
1

I think I might misunderstand EfficientDet-lite output. I’ve been working with the salad detector example and have been able to train it to recognize my object. However it also detects a bunch of other things and classifies them as my object. For example if I show it a picture of my object and a person, it will outline the person, their hand, and my object and all are classified as my object. I would have thought it would say they were a person.

Now I’m starting to think I’m wrong. And that since I trained it with only my object, then that is the only object it understands. I’m not sure if I’m right about this though. So then I went and just ran the Salad detector example from the beginning with no changes. And I took that model and ran it against my person + my object photo. This time in recognized only the person’s shirt and the only classes shown were from the parts of a salad.

So I guess maybe I misunderstand how EfficientDet-lite is supposed to work? Maybe I need to train it with a lot more data about my object so I don’t get so many false positives?

Thanks!

3

Hi @Eric_Steimle ,

Yes, Make sure your training data is diverse and representative of the scenarios where you expect the model to perform well. Include various backgrounds, lighting conditions, and orientations for your object. This can help the model generalize better.

Remember that the effectiveness of the model depends on the quality and diversity of your training data. Experiment with adjustments in training data and model parameters to achieve the desired balance between precision and recall for your specific use case and training object detection models is an iterative process, and it might take some adjustments and experimentation to achieve the desired results.

Thanks.