Skip to content

DenoiSeg Performance Comparison with Other Semi Supervised Baselines

Mangal Prakash edited this page May 24, 2020 · 6 revisions

Upon reviewers' request, we have performed additional performance comparisons with state-of-the-art semi-supervised methods, namely Finetune and Finetune Sequential schemes which were reported to outperform U-Net and the sequential schemes mentioned in our manuscript (The baseline and sequential schemes in the graphs below are the same as reported in our manuscript). Our results show that we outperform these semi-supervised baselines as well by a large margin in terms of both Average Precision (AP) and SEG scores for various noise levels, especially when only limited ground truth labels are used for training the segmentation task. For many more qualitative and quantitative results, please refer to the results section.

AP (left) and SEG (right) scores for DSB data corrupted with gaussian noise of zero mean and standard deviation 20

DSBn20

AP (left) and SEG (right) scores for DSB data corrupted with gaussian noise of zero mean and standard deviation 10

DSBn10 Figure: On the x-axis, we show the number of ground truth segmentation images available for training any particular scheme. Our DenoiSeg method with default alpha = 0.5 as well as our DenoiSeg scheme with best alpha outperform all other baselines for low number of annotations. Please refer to our paper to find out details on how the best alpha is computed.