"Feature explorer" graph for object detection

I noticeed points representing some images in my object deetection project are at the position 0,0 in the 3D plot. Is this because of a bug or has special meaning ? The same photos in a previous version with less categories were not in the (0,0) position.

Also, I am interested to know your thoughts about how far I can get benifit from the “Feature explorer” to reduce the anomolies in my object detection dataset. I know that this feature is based on UMAP algorithm which reduces the data dimentionality, and the orginal paper showed how MNIST was nicly clustered to 10 clusters (image bellow). This should be very similar to any calssification problem. But, in object detection problem (which part of it is a classification), each image has many objects inside not only one so could that be really nicly clusstered using UMAP or even close?


Hi @yahyatawil,

Interesting about some samples marked as (0, 0, 0) in the feature explorer. I’ll post this to the devs to see if it’s a bug.

You can apply UMAP to each object instance in the various dataset images. So, while a single image might have multiple classes of objects, each object can be mapped to the 3-axis feature explorer. That’s why the individual dots show you “fence,” “fire_hydrant,” etc. versus “image_01,” “image_02,” etc.

Thanks Shawan for your reply. I’ll be waiting your developers reposnse.

Regarding the UMAP for object detection dataset, do you mean to apply the UMAP with only one class enabled at a time? and repeat that accross all classes?

Also can you confirm that the UMAP is applied to the complete image or for the boudning box instance window only ?