PSPNet-Pytorch

This Repo contains an implemetation of "Pyramid Scene Parsing Network" in Pytorch. Pretrained weights are converted from Official Caffe Repo. The performance is: [acc 0.95196], [acc_cls 0.83369], [mean_iu 0.75547], [fwavacc 0.91172] on Validation set of CityScapes Dataset.

Installation

1. Install Anaconda3 from Here
2. Create a Conda Environment:

username@PC:~$ conda env create -f environment.yml

3. Download Cityscapes Dataset from Here
4. Download Caffe Pretrained from Here and put it in Caffe-PSPNet folder
5. Run the code:

username@PC:~$ python main.py

Qualitative

Note

Weight conversion from Caffe to Pytorch is modified from pytorch-semseg.
Preprocessing and loss function is modified from pytorch-semantic-segmentation.
Since the differences are significant, I decided not to add a branch to any of the the above project.