FOMO: detecting very small objects

Hello there,
I was wondering, when looking at the Keras script (in the expert interface) whether and how it is possible to tweak the FOMO object detection models to also perform well with very small objects. I couldn’t find any obvious parameters to tweak. Our objects are around 100x100 px big and taken from a full-sized openMV H7+ image (that is, 2492x1944). Our challenge is similar to the successful bee detection example that is on the FOMO documentation page (can’t post - new user).

Some of my models that should detect small flowers and fruits on trees perform badly with overall accuracies around 30%. We have a relatively small dataset of 40 images, but there are plenty of labels (an estimated minimum of 400, so 10 per image) and I was thinking that this should suffice. Any suggestions? @shawn_edgeimpulse

Here is an example of a labeled image:

Hi @darrask,

you can tune the spatial reduction using the Keras expert mode. See more information in our documentation here.


1 Like

Thanks! I tried all possible (3-16, in steps of two) spatial reduction values but did not achieve substantially better performance. Will try object weighting and doubling the training image dataset.

1 Like