Can FOMO access the BACKGROUND Value?

#1

I ask the stupid questions so other people don’t have to!

Can an EdgeImpulse Vision FOMO model result access the BACKGROUND value? I can get all the other values using

    for (size_t ix = 0; ix < EI_CLASSIFIER_OBJECT_DETECTION_COUNT; ix++) 
    {
         auto bb   result.bounding_boxes[ix]
         // you can now use, be aware of differing variable types  
         // bb.label,   bb.value,    bb.x,     bb.y,     bb.width,     bb.height
    }

image
image609×573 33.1 KB

If there is a variable for the BACKGROUND, does it also have an x, y, width, height or does that not make any sense? Not sure if I would ever use the data, I just want to know if it is available.

#2

Hello @Rocksetta,

That’s a good question but I am afraid I don’t know the answer…
Maybe @matkelcey can answer that.

Seeing how FOMO is built, I believe each “cell” should have its class classification between your objects and background and they could be accessible in a way (probably with some tweaks in our SDK). Also we perform a post processing to avoid having too many classes that are too close to each other and that actually represent the same object. So you would not be able to get this post-processing if you want to access the predictions for each cell.

Regards,

Louis

1 Like
#3

Hi @Rocksetta if you have an EON model and call trained_model_invoke() you’ll get an output tensor with the heatmap w/ n-number of classifications per cell (where n=class count + 1 background class); but you’ll need to do all the wiring yourself - see how we do it in the SDK here inferencing-sdk-cpp/ei_run_classifier.h at 144730e047d370526ca4b7ecdc247503552f3020 · edgeimpulse/inferencing-sdk-cpp · GitHub

2 Likes