search

lxtGH / DecoupleSegNets

[ECCV-2020]: Improving Semantic Segmentation via Decoupled Body and Edge Supervision
370 stars 36 forks source link

Does anyone get the 80.7% mIoU with Resnet101 backbone for training the Deeplabv3+? #22

Closed lqxisok closed 3 years ago

lqxisok commented 3 years ago

We reimplement the Deeplabv3+ model and run with the following configuration. 2da9e08e774cb40bc8d45c9ddb2651e

But the final mIoU is just 78.9%. So my question is how to get the 80.7%mIoU performance reported in the DecoupleNet paper? We want to reimplement the outstanding result. Due the the GPU memory limitation, we use AMP training to save the memory requirement. Does it influence the performance mostly? Hope the author or many good brothers help me solve the question.

lqxisok commented 3 years ago

I found that is my implement problem. Refer to mmsegmentation I got the right performance