Question about calculating uncertainty of non-bayesian model when training

[lisosia]

Hi, Thank you for your great work.

When training for non-bayesian model,
it seems that certainty map is calculated by softmax(logit):

UMA/train_bayesian_adapt.py

Lines 359 to 362 in b72dad9

	predict = cv2.resize(predict, img_size, interpolation = cv2.INTER_CUBIC)
	cv2.imwrite("%s/%s_p.png" % (saveDir, filename[:-4]), predict)
	predict_seg = cv2.resize(predict_seg, img_size, interpolation = cv2.INTER_CUBIC)
	cv2.imwrite("%s/%s_s.png" % (saveDir, filename[:-4]), predict_seg)

Is it more reasonable to use abs(softmax(logit) - 0.5) ?
It seems that Figure2(c) use something like the latter equation to calc (un)certainty map.

Thanks.

[cai-mj]

Hi, Thank you for your great work.

When training for non-bayesian model,
it seems that certainty map is calculated by softmax(logit):

UMA/train_bayesian_adapt.py

Lines 359 to 362 in b72dad9

	predict = cv2.resize(predict, img_size, interpolation = cv2.INTER_CUBIC)
	cv2.imwrite("%s/%s_p.png" % (saveDir, filename[:-4]), predict)
	predict_seg = cv2.resize(predict_seg, img_size, interpolation = cv2.INTER_CUBIC)
	cv2.imwrite("%s/%s_s.png" % (saveDir, filename[:-4]), predict_seg)

Is it more reasonable to use abs(softmax(logit) - 0.5) ?
It seems that Figure2(c) use something like the latter equation to calc (un)certainty map.

Thanks.

Hi, my code doesn't calculate uncertainty map for non-bayesian model. The comment before that four lines of code is mistakenly written (sorry).
The Figure 2(c) in the paper was indeed visualized based on uncertainty = 1 - abs(softmax(logit) - 0.5)*2 for demonstration.

[lisosia]

Thanks for your clarification.
Is the equation uncertainty = 1 - abs(softmax(logit) - 0.5)*2 also used for the experimant of Table1 (CNN+uma) ?