NMS start suppressing crashes

[usmanqureshi94]

Describe the bug
I have a small tile, for which startdist detects about a 0.7million points or n_polys. In nms start suppressing step, the program crashes after hanging the system for a while. I understand it's very unusual to have so many points in such a small tile but how can I overcome this issue of stardist crashing.

To reproduce
I've attached the image to reproduce the issue with a pre-trained model.

[usmanqureshi94]

[uschmidt83]

Sorry for the late reply.

I've attached the image to reproduce the issue with a pre-trained model.

Can you please be more specific how to reproduce this issue?

[usmanqureshi94]

No worries, if you download the above image and just call stardist's predict method. it finds about 700k potential points at first and gets stuck in nms loop.

[uschmidt83]

I guess you mean the predict_instances method that includes the NMS loop. Which model are you using?

[usmanqureshi94]

I tested with both pretrained stardist 2d model and custom model 2d model. Whenever points are more than 500k the execution gets stuck in the NMS loop. I tested with 8 core and 32 core CPU but the execution never terminates.

[uschmidt83]

Your image is mostly just white, hence not suitable for any of our pre-trained models (which expect a black background).

I did the prediction with our 2D_versatile_fluo model, and the probability image is almost 0 everyhere, i.e. no objects would've been predicted. I strongly suspect you didn't normalize your image before prediction (e.g. try model.predict_instances(img/255)).

[usmanqureshi94]

thanks this resolved the issue. I was using csbdeep's normalize function as per example notebook
img_gray_slice = normalize(img_gray_slice, 1, 99.8, axis=cfg.AXIS_NORM)