Introduction

This code is based on Deeplab v2, which I merged some other features:

• Fixed the batch_norm_layer's bug by replacing with caffe-master's batch_norm_layer
• Merge crop_layer from caffe-master for FCN support
• Merge multi_stage_meanfield layer from https://github.com/bittnt/caffe/tree/crfrnn for CRFasRNN merge
• Modify image_seg_data_layer.cpp to support class segmentation label index > 256. if set label_span: RG in prototxt,Index=R/10*256+G.
• Merge LRCN support with coco caption example included.
• Merge PSPNet support by merging bn_layer from PSPNet while without MPI-support for multi-batch train currenttly.

Tips

• How to load source image & label image for training?

  Use python interface as FCN's or deeplab's filelist format. examples/indoor_scene_parsing/scripts/image_seg_data_layer.py maybe help too.

• How to fetch image to segmentation models with any size?

  Ref to FCN's python script. examples/indoor_scene_parsing/scripts/predict.py maybe help too.

• How to display result pretty?

  examples/indoor_scene_parsing/scripts/viz_seg.py maybe help.

• How to use CRFasRNN?

  examples/indoor_scene_parsing/test_pspnet50_ade20k64cls_v2.prototxt is an example. CRFasRNN's README may help too, but some differs,

   layer {
     name: "splitting"
     type: "Split"
     bottom: "conv6_interp"
     top: "unary"
     top: "Q0"
   }
   layer {
     name: "crf_inf"
     type: "MultiStageMeanfield"
     bottom: "unary"
     bottom: "Q0"
     bottom: "data"
     top: "crf_inf"
     param { lr_mult: 10000 }
     param { lr_mult: 10000 }
     param { lr_mult: 1000 }
     multi_stage_meanfield_param {
   	num_iterations: 3
   	compatibility_mode: POTTS
   	threshold: 2
   	theta_alpha: 160
   	theta_beta: 3
   	theta_gamma: 3
   	spatial_filter_weights_str: "3 3 2"    # Remain unsetting value = 3
   	bilateral_filter_weights_str: "5 3 4"  # Remain unsetting value = 4
     }
   }
  

Install

The same to Deeplab v2.