Reproducing-PSMNet

Overview

This is a reproduction of the CVPR 2018 paper "Pyramid Stereo Matching Network". This repository contains code to calculate end-point error on Scene Flow dataset and to visualize the training and validation loss.

Before starting to work on the code read our blog post on the same.

We have provided files that we made to calculate the end point error and visualizing the training loss. To know more please refer to this repository .

File description

1. reproduce.py - Calculates the end point error

2. finetune.py - This file is same as from the original repository with additional code to store losses.

3. training_and_validation_plot.py - Plots the training and validation loss.

Calculating the end point error

• Step 1: Clone the PSMNet repository from the above link.

• Step 2: Clone this repository.

• Step 3: Copy the files from this repository and paste it under the directory /your_path/PSMNet-master/

• Step 4: Specify the path to your trained or pretrained model in the reproduce.py and also to the folder containing Scene FLow test dataset.

• Step 5: Run the code to obtain the end point error of your model.

Visualzing the training and test loss

• Run the script finetune.py that you copy pasted from this repository. This will save a file containing all losses that occurred during finetuning.

• Run the script training_and_validation_plot.py. You will see your training and validation loss printed to the console.