search

dbolya / yolact

A simple, fully convolutional model for real-time instance segmentation.
MIT License
5k stars 1.33k forks source link

Probability Maps #713

Open MiguelAngeloMartinsRibeiro opened 2 years ago

MiguelAngeloMartinsRibeiro commented 2 years ago

Hi @dbolya Is it possible to get the probability maps, i.e, instead of binary masks get a matrix with the probability of each pixel?

Thank you for your time

MiguelAngeloMartinsRibeiro commented 2 years ago

The softmax output would be useful as well

telemetrieTP23 commented 2 years ago

You are right the probability map Output would really be nice for Heatmap Visualisation of a specific class. Convert the Modell to ONNX and visualize it with https://netron.app/ So you can find the specific layers you need. You can also use pre converted ONNX yolact/yolactEdge Modells from https://github.com/PINTO0309/PINTO_model_zoo

If you use Python then you can use this as starting point for Postprocessing Layers: https://github.com/openvinotoolkit/open_model_zoo/tree/master/demos/instance_segmentation_demo/python

For c++ use this: https://github.com/PINTO0309/PINTO_model_zoo/blob/main/085_Yolact_Edge/30_Full_Converted_mbnv2_550x550/yolact_test.cpp