Monocular Neural Image-based Rendering with Continuous View Control

Unofficial implementation of 'Monocular Neural Image-based Rendering with Continuous View Control' using mxnet gluon. This model can generate novel views of objects from only one view, with fine-grained control over the virtual viewpoints. This repository contain some code from the original implementation and some parts of the encoder-decoder changed.

Paper: https://arxiv.org/abs/1901.01880

Official implementation: https://github.com/xuchen-ethz/continuous_view_synthesis

Reproducing Results

Installation

Install all packages with pip, run

pip install -r requirements.txt

mxnet 2.0.0 can be found here https://dist.mxnet.io/python/all

Training

-- First download the dataset (I only implemented the code for 'car' or 'chair' dataset) from Google Drive
-- Create new folder and extract the dataset to the new folder
-- Train the model with

python train.py --dataset_path 'new_folder_path'

If the new folder is 'C:\Datasets\continuous_view_synthesis_dataset' the command is:

python train.py --dataset_path 'C:\Datasets\continuous_view_synthesis_dataset\'