search

wuhuikai / FastFCN

FastFCN: Rethinking Dilated Convolution in the Backbone for Semantic Segmentation.
http://wuhuikai.me/FastFCNProject
Other
838 stars 148 forks source link

Comparison with deeplab v3 #53

Closed siju-samuel closed 4 years ago

siju-samuel commented 4 years ago

deeplabv3 resnet50 Model size: 98MB Parameters: 39,638,869

deeplabv3 resnet50 with JPU Model size: 489MB Parameters: 63,931,934

https://github.com/wuhuikai/FastFCN/blob/097e7130f77a15eeca12f37a0afcf2f8f7f90439/encoding/models/deeplabv3.py#L16

If you see the above link, we can see the architecture with JPU contains [resnet50(dialations removed) + JPU + ASPP_Module + FCNHead] whereas normal deeplabv3 consists of [resnet50 + ASPP_Module] This difference explains the memory/parameter size

Please correct me if im wrong. Im not able to understand how this will improve the memory footprint & speed of the network in comparison with deeplabv3 with/without jpu.

wuhuikai commented 4 years ago
  1. The model size of DeepLab-V3 is 338MB[resnet50+ASPP_Module+FCNHead].
  2. FCNHead is used for deep supervision.
  3. The output stride of DeepLab-V3 is 8 while that of DeepLab-V3-JPU is 32, thus DeepLab-V3-JPU can improve memory footprint and speed.