FOMO using center point detection and not bounding boxes

So basically I’m training a fomo model, but i want to do the center point detection and not the default bounding boxes as they’re leading to my model being overfitted. Can I change this to center point detection? I’m just on a student plan here.

Thanks in advance.

Hi @meow33 - I don’t quite follow your question. Perhaps you can elaborate?

When training a FOMO model, the training data uses bounding boxes but the inference results return centroids. This can’t be changed for FOMO.

If your model is overfitting, you can look at increasing the size and diversity of your training data set, using data augmentation, using fewer training epochs, implementing early stopping conditions (in expert mode), etc.

1 Like

Hello Brian

Firstly, thank you for clarifying this. I appreciate this. My model that I’m training is running on FOMO MobileNet V2 0.35, and I’m doing it for object detection and classification. My dataset contains images that may not be easily aligned into a bounding box when labeling, and hence, I sometimes get varied results. Also, after a few tries with modifications and higher epochs, I learnt that my model seems to be overfitted, so I reduced the number of epochs and also enabled the data augmentation option. But another problem arose that while in testing, the images that the model classifies wrongly have an F1 score of 0. That’s right, it somehow manages to generate features and classify other images, but for certain images, it fails to do this. This is also creating problems in deployment. I’d appreciate your help here immensely.