Closed seekingdeep closed 3 years ago
Hi @seekingdeep Thanks for looking into EmbedSeg.
Regarding the second question about comparisons with DenoiSeg, since there the model is trained for a semantic task (i.e. trained to predict the pixel class) and addresses instance segmentation by extracting connected components during post processing, this would generally under-perform methods on instance segmentation metrics (such as AP and SEG) where the model is trained directly for instance segmentation.
For example: AP scores for IoU threshold = 0.5 for alpha best configuration using all images (3800) as quoted by DenoiSeg is 0.801±0.001 (in comparison, instance segmentation methods generally perform at ~0.86 on the same dataset!)
The benefits of DenoiSeg and other methods addressing semantic segmentation on the other hand is a simplified architecture, more intuitive methodology and more robust training. Also, DenoiSeg should perform equally well as instance segmentation approaches on semantic segmentation metrics (since this is what the model is trained for) such as computing Jaccard Index between the predicted foreground and the ground truth foreground.
Regarding the first question, we re-used and adapted code by Neven et al, 2019 and so wanted to use the same license file as used by them!
Thanks
embedseg seems promising,