search

open-mmlab / mmsegmentation

OpenMMLab Semantic Segmentation Toolbox and Benchmark.
https://mmsegmentation.readthedocs.io/en/main/
Apache License 2.0
8.02k stars 2.58k forks source link

Confidence threshold #3528

Open 0ttwhy4 opened 8 months ago

0ttwhy4 commented 8 months ago

I am trying to use pretrained models in mmseg to obtain a certain class of objects in images,which will serve as ground truth in my own model'training,and thus I want the result to be as accurate as possible by screening out pixels with low confidence.So how should I do it?:|

0ttwhy4 commented 8 months ago

Never mind.Already solved : )

1wang11lijian1 commented 7 months ago

@0ttwhy4 Hello, how do you get the confidence of the segmentation results to filter out some predictions that do not meet the conditions.

changwsh12 commented 7 months ago

Assume there are 2 classes, the following can save the probabilities of each class and the classification result in csv files (you can also directly use the intermediate variables to do other things): model = init_model(cfg, checkpoint_file, 'cuda:0') result = inference_model(model, image)

save the probabilities of the first class

np.savetxt('logits0.csv', result.seg_logits.data[0].to('cpu').numpy(), delimiter=',')

save the probabilities of the second class

np.savetxt('logits1.csv', result.seg_logits.data[1].to('cpu').numpy(), delimiter=',')

save the class number

np.savetxt('pred.csv', result.pred_sem_seg.data[0].to('cpu').numpy(), delimiter=',')

1wang11lijian1 commented 7 months ago

@changwsh12 Ok, got it. Thank you for your answer