eqiihuu / residual_ResNet

An improvement to SegNet using residual unit

SegNet with Residual Unit

Efficient Multi-Scale Residual Network for Image Classification and Semantic Segmentation. We use residual unit in the sub-structure of SegNet to improve the performance.

Datasets

• CamVid dataset is provided in the data/CamVid folder. If you are using this dataset, please cite the following paper

  i) Brostow, Fauqueur, Cipolla, Segmentation and Recognition Using Structure from Motion Point Clouds, ECCV 2008
  ii) Brostow, Fauqueur, Cipolla, Semantic Object Classes in Video: A High-Definition Ground Truth Database, Pattern Recognition Letters

• Download the PASCAL VOC 2012 dataset using below link

wget http://cvlab.postech.ac.kr/research/deconvnet/data/VOC2012_SEG_AUG.tar.gz

Training Semantic Segmentation Models

You can train the network as:

• Training M-RiR using GPU-0 on the CamVid dataset

      CUDA_VISIBLE_DEVICES=0 th main.lua -dataset cv -imHeight 384 -imWidth 480 -modelType 2 -lr 0.0001 -d 10 -de 300 -optimizer adam -maxEpoch 100
  

• Training M-Plain using GPU-0 and GPU-1 on the CAMVID dataset

    CUDA_VISIBLE_DEVICES=0,1 th main.lua -dataset cv -imHeight 384 -imWidth 480 -modelType 1 -lr 0.0001 -d 10 -de 300 -optimizer adam -maxEpoch 100
  

• Training M-Hyper on the CAMVID dataset

    CUDA_VISIBLE_DEVICES=1,2 using GPU-1 and GPU-2 th main.lua -dataset cv -imHeight 384 -imWidth 480 -modelType 3 -lr 0.0001 -d 10 -de 300 -optimizer adam -maxEpoch 100
  

Training Image Classification Models

• For training on the ImageNet and the Cifar datasets, please use the scripts provided by FaceBook AI Research
• Once you get the source code from FaceBook AI Research github repository, please replace the content of resnet.lua by mresnet_class.lua
• Follow the instructions mentioned on FaceBook AI Research Github page for training.
• For doing depth related studies, please follow below command

To train MResNet on Cifar10 dataset with a depth of 11

th main.lua -dataset cifar10 -nGPU 2 -batchSize 128 -depth 1

To train MResNet on Cifar100 dataset with a depth of 20

th main.lua -dataset cifar100 -nGPU 2 -batchSize 128 -depth 2

License

This software is released under a creative commons license which allows for personal and research use only. For a commercial license please contact the authors. You can view a license summary here.