Output

Phase 1: Structure From Motion (SfM)

Run the wrapper.py file in phase 1 folder.

Phase 2: NeRF

This project trains a Neural Radiance Field (NeRF) model on synthetic datasets. The implementation leverages PyTorch for model training and evaluation.

Environment Setup

To run this project, ensure you have Python 3.8+ and pip installed on your system. It's recommended to use a virtual environment:

python3 -m venv nerf-env
source nerf-env/bin/activate

Install the required packages:

pip install torch torchvision opencv-python tqdm numpy matplotlib tensorboard

Project Structure

• train.py: Main script for training the NeRF model.
• test.py: Script for evaluating the trained model.
• nerfnetwork.py: Defines the NeRF model architecture.
• loader.py: Contains code for loading and preprocessing the dataset.
• render.py: Functions for rendering images using the trained NeRF model.
• utils.py: Utility functions used across the project.
• psnr_ssim.py: Script for calculating PSNR and SSIM metrics.
• /checkpoints: Directory to save model checkpoints.
• /runs: Directory where TensorBoard logs are saved.
• /output: Directory for saving rendered images and other outputs.

Training

To train the NeRF model, use the following command:

python3 train.py --batch_size 1024 --epochs 10 --learning_rate 1e-4 --width 256 --depth 8

Command-Line Arguments

• --batch_size: The size of the batch. Default is 1024.
• --epochs: Number of training epochs. Default is 10.
• --learning_rate: Learning rate for the optimizer. Default is 1e-4.
• --width: Width of the layers in the NeRF model. Default is 256.
• --depth: Depth (number of layers) of the NeRF model. Default is 8.

Evaluation

To evalua

python test.py

Make sur

e to modify test.py to load the correct model checkpoint and specify the correct dataset paths.

Visualization

Training progress and loss can be visualized using TensorBoard:

bash
tensorboard --logdir runs

Open the provided URL in a web browser to view the training logs.

Additional Notes

• Ensure the dataset is correctly placed as per the path specified in loader.py.
• The checkpoints and output directories will be created automatically to store model checkpoints and output images, respectively. If not, make the directory in the current working directory.