I am trying to detect objects that are not in the standard COCO dataset classes.
I presume if I download the model, and discard the weights (I think ideally not all of them but only the last layers?), it could be retrained with new classes.
I have searched quite a bit but I think most arent really re-trainable and I am quite confused after days searching. For example.
Do you have any example or could you tell me if there is any model that could be re trained for a specific class in the way I described above?
Hi @Mah_Neh, you can retrain the object detection models on your custom dataset that are present in the model garden using the object detection api. For more details please refer to this document. Thank You.
I do not like those overloaded approaches. I am looking for this:
- The NN in code or a json config file
- The weights
The I only need to resize the dataset, load a layers model and fit. It seems straightforward. Why isnt it done like that @Kiran_Sai_Ramineni ?
A better way is through
keras.applications imho. This is a lot more like I imagine it, but no many models there.
KerasCV has object detection models such as RetinaNet and YOLOv8! See our new guide!