search

0bserver07 / One-Hundred-Layers-Tiramisu

Keras Implementation of The One Hundred Layers Tiramisu: Fully Convolutional DenseNets for Semantic Segmentation by (Simon Jégou, Michal Drozdzal, David Vazquez, Adriana Romero, Yoshua Bengio)
https://arxiv.org/abs/1611.09326
MIT License
197 stars 54 forks source link

About the reslut? #11

Open youngfly11 opened 6 years ago

youngfly11 commented 6 years ago

Hi, You have done an awesome work. But i have a question to ask you. Have you reproduced the same result reported in the original paper, Mean_IoU=66.9%,acc=91.5 ? If you have reproduce the same result, please tell me what kind of trick you used in your network?

Thank you in advance

0bserver07 commented 6 years ago

hey @youngfly11 I haven't been able to reproduce the same exact results :/, though I have a new branch that I will push in a week or so, which improves the results on that dataset.

Faur commented 6 years ago

Looking forward to it!

pjohh commented 6 years ago

Hi @0bserver07 do you have any updates regarding the results (and the repo)?

0bserver07 commented 6 years ago

Ow wow! I had totally forgotten about the fact that I never pushed the updates from that time. Yes, I figured out the issue, it was partially my way of writing the model. (in span of 2weeks, will push a whole new verion)

pjohh commented 6 years ago

Hey @0bserver07 thanks for the update! I implemented it myself yesterday in keras and figured out, that your DenseBlock is implemented incorrectly and concatenations/skip connections are missing... But you already know I guess. I'm looking forward to compare my solution with yours, since the tiramisu paper is very ambiguous

ohernpaul commented 6 years ago

@pjohh I am about to go through the same process of fixing the skip connections. Could you make a PR or send me the fix to your model file? Thanks!

PhilippKopp commented 6 years ago

hey @pjohh as @ohernpaul I am also experiencing the same issues with the skip connections. A PR with your fix would help a lot. Thank you!

dongzhuoyao commented 6 years ago

@0bserver07 could you tell me you best result?

dongzhuoyao commented 6 years ago

I think the experiment data is wrong: why the gap between the result of segnet and bayesian segnet is so huge?? this makes this paper result incredible 💢