search

jiwoon-ahn / irn

Weakly Supervised Learning of Instance Segmentation with Inter-pixel Relations, CVPR 2019 (Oral)
MIT License
519 stars 100 forks source link

deeplab-v2 and CRF #24

Closed arnike closed 4 years ago

arnike commented 4 years ago

Hi Jiwoon, the original deeplab-v2 with VGG16 and ResNet-101 have somewhat different architectures (e.g. design of ASPP module). I was wondering, in your implementation with ResNet-50, did you use ResNet-101 as the reference, or the VGG-based one? Also, from A.2 it seems that you used CRF to compute the upper bound. Did you also use CRF after fully-supervised training on the pseudo labels? Thanks in advance, Nikita

jiwoon-ahn commented 4 years ago

Hi @arnike, Sorry for the late reply, 1) We set ResNet-101 as the reference, and reduced its depth to ResNet-50. 2) Yes, we did so because the original DeepLab v2 employed dCRF.