Getting bad F1 score even though adding more images

Hi, I got the F1 score of 80 % initially using the FOMO model. However, I’m getting a bad F1 score of 59 % after adding more images for training. I have tried to change the settings but I can’t even figure out what’s the problem causing it. So, how can I increase F1 score? My project ID: 153503

Thank you !

Hi @sio_yx,

The issue you are running into is a limitation with FOMO itself. If you have multiple representation grid cells next to each other, FOMO will pick the highest-scoring grid cell to classify/localize as your object. I highly recommend taking a look at my talk here to learn more about the limitations of FOMO: tinyML Talks: Constrained Object Detection on Microcontrollers with FOMO - YouTube

For example, FOMO is going to struggle with the image below, as all of the objects are next to each other. As a result, the only 1 or 2 cells will be counted. This hurts your F1 score, even though FOMO is working as intended.


Thank a lot @shawn_edgeimpulse

1 Like