I wanted to compare how our thresholding compared with other automatic thresholding techniques, like Otsu. I created a file: evaluate_auto.py that compares our sweeping with other ones. Feel free to edit it however you see fit. I tested this out on the max model on fold 0. Here are the results:
====================================================================================================================
Method Mode thr_bld thr_veg thr_wat iou_bld iou_tree iou_wat RMSE_bH RMSE_vH Score
--------------------------------------------------------------------------------------------------------------------
sweep:base(0.5) global 0.5000 0.5000 0.5000 0.5290 0.7725 0.5379 1.6386 2.9093 0.5259
sweep:global(0.650) global 0.6500 0.6500 0.6500 0.5358 0.7699 0.5441 1.6386 2.9093 0.5281
*sweep:per-class global 0.6750 0.5500 0.9000 0.5358 0.7733 0.5530 1.6386 2.9093 0.5300
otsu global 0.4121 0.5215 0.4746 0.5234 0.7730 0.5359 1.6386 2.9093 0.5242
otsu per-image varies varies varies 0.5153 0.7703 0.4823 1.6386 2.9093 0.5138
yen global 0.0293 0.3536 0.0020 0.4423 0.7604 0.3394 1.6386 2.9093 0.4726
yen per-image varies varies varies 0.3750 0.6705 0.3402 1.6386 2.9093 0.4424
li global 0.1104 0.3487 0.1114 0.4849 0.7597 0.5015 1.6386 2.9093 0.5075
li per-image varies varies varies 0.4765 0.7551 0.4541 1.6386 2.9093 0.4976
triangle global 0.0098 0.9668 0.0059 0.3993 0.6024 0.4029 1.6386 2.9093 0.4477
triangle per-image varies varies varies 0.3851 0.6352 0.3859 1.6386 2.9093 0.4465
isodata global 0.4121 0.5215 0.4746 0.5234 0.7730 0.5359 1.6386 2.9093 0.5242
isodata per-image varies varies varies 0.5156 0.7703 0.4828 1.6386 2.9093 0.5140
pct90 global 0.0044 0.9980 0.0004 0.3618 0.2098 0.1834 1.6386 2.9093 0.3465
pct90 per-image varies varies varies 0.2146 0.2769 0.1113 1.6386 2.9093 0.3089
--------------------------------------------------------------------------------------------------------------------
Our sweep still performs well, but there might be other automatic thresholding techniques to consider, like Otsu and Isodata that may help our performance.
I wanted to compare how our thresholding compared with other automatic thresholding techniques, like Otsu. I created a file: evaluate_auto.py that compares our sweeping with other ones. Feel free to edit it however you see fit. I tested this out on the max model on fold 0. Here are the results:
Our sweep still performs well, but there might be other automatic thresholding techniques to consider, like Otsu and Isodata that may help our performance.