search

wasidennis / AdaptSegNet

Learning to Adapt Structured Output Space for Semantic Segmentation, CVPR 2018 (spotlight)
850 stars 203 forks source link

About baseline result with res101-deeplabv2 #16

Closed shanyuhu closed 6 years ago

shanyuhu commented 6 years ago

Hi @wasidennis, I use your provided res101 backbone and train deeplab v2 on GTA as your code, Then I evaluate the model on cityscapes, but I only got 31.1 which is far from your baseline result of 36.6. Did I miss anything? Looking forward to your reply.

shanyuhu commented 6 years ago

@hfslyc

shanyuhu commented 6 years ago

Another question is that why the oracle resutls of res101 in Table 2 and Table 5 are different? one is 65.1, and the other is 71.7.

tuanhungvu commented 6 years ago

Hi @shanyuhu , I got the same result. I wonder if you solved the problem?

shanyuhu commented 6 years ago

@tuanhungvu my final resule is around 34%~35%, still lower than the reported number in the paper

tuanhungvu commented 6 years ago

@shanyuhu I see. Did you change anything to get this result (e.g. init LR, params scheduler, optimizer...) ?

tarun005 commented 4 years ago

@tuanhungvu @shanyuhu I wonder how you solved your problem. I am getting ~29% on baseline ResNet-101.

adricarda commented 4 years ago

Did anyone solve the baseline problem? I only get about 29% too.