Linknet is a new architecture for semantic image segmentation by the authors of E-Net. Details of the paper are not really clear from this blog. They have released the code on github, but without a strong understanding of their results with different architectures, I am not really sure what to try for best results.
Can anyone help me out here? Has anyone tried playing out with skip connections of E-net and give me insights into the results?
Probably easiest for wait for the paper. I chatted to the author about it briefly. I suspect it’ll be somewhat like 100 layer tiramisu, but with the enet rather than densenet architecture as starting point.