#PspNet

• Author: Manish Soni Email: [email protected]
• License: license

PspNet is in Caffe

This repo attempts to reproduce this amazing work by Hengshuang Zhao : PspNet

Requirements

• [Cuda-8.0]
• [Opencv-3.1.0]
• [Minimum 8Gb Gpu RAM]

Downloads

- Download original Pspnet trained models and put them in folder pre_trained_model
  $cd PspNet-Caffe_Root
  $wget https://drive.google.com/file/d/0BzaU285cX7TCN1R3QnUwQ0hoMTA/view
- Download Pspnet model trained on Amazon robotics Challenge dataset
  $cd PspNet-Caffe_Root
  $wget https://drive.google.com/file/d/0BzaU285cX7TCN1R3QnUwQ0hoMTA/view

How To

1. Firt clone the repository 
-  $git clone https://github.com/ManishSoni1908/PspNet-Caffe.git
2. Go to  ./caffe-pspnet  folder, make build folder and run cmake ..  and then make - j[Number of cores].
3. Go to  ./pspnet_test  folder, open CMakeLists.txt and mention caffe build path.
4. Go to  ./pspnet_test  folder make build folder and run
- $cmake .. 
- $make all.

Training

1. Go to  ./utils/prototxt_training/solver_PSP.prototxt  Mention the network prototxt, snapshot path and max_iteration.
2. Go to  ./utils/prototxt_training/original_pspnet.prototxt  Mention the NumClasses in conv6_apc17 and conv6_1_apc17 layer.
3. Go to  ./caffe-pspnet/build/tools` Run 
- $./caffe train -gpu all -solver [Path to solver.prototxt] -weights [path to pre-tarined model]

Testing

1. Go to `./pspnet_test/build` Run
- $./test [path of testing prototxt] [path of trained model] [path of test image directory]
2. All done

##Results