Switch branches/tags
Nothing to show
Find file History
Fetching latest commit…
Cannot retrieve the latest commit at this time.
Type Name Latest commit message Commit time
Failed to load latest commit information.



  • RefineNet was originally proposed in: [1] Guosheng Lin, Anton Milan, Chunhua Shen, Ian Reid RefineNet: Multi-Path Refinement Networks for High-Resolution Semantic Segmentatiion. arXiv:1611.06612

  • The RefineNet implemented in this module is based on [1]

  • Since RefineNet relies on ResNet, we use ResNet-50 for this implementation

    • resnet_v2.py and resnet_utils.py are copied from github repo tensorflow/models/research/slim/nets
    • resnet_v2_50 in resnet_v2.py is used to create the 50-layer ResNet
  • The key differences between this implementation and the one proposed in [1]:

    • [1] uses ResNet pretrained on ImageNet recognition tasks, while this implementaton is trained end-to-end
    • [1] uses 512 filters for each conv layer of RefineNet-4 block, while this implementation uses 256 instead, to keep it consistent with the remaining RefineNet blocks
  • This implementation only supports input images that are 512x512x3. Other sizes might not work.

  • tensorflow 1.5.0 or above is required. Using lower versions of tensorflow may generate "incompatible dimension" errors.

  • pretrain_resnet.py can be used to pretrain ResNet50 defined by slim's resnet_v2_50 in resnet_v2.py. wrangle_tiny_imagenet.py is used to prepare the raw Tiny ImageNet dataset to the file structure that Keras image generator supports.