https://github.com/alessiopittiglio/dl-project

A TensorFlow implementation of Neural Radiance Fields (NeRF) for synthesizing novel views of complex 3D scenes from a sparse set of 2D images.
https://github.com/alessiopittiglio/dl-project

computer-vision nerf novel-view-synthesis tensorflow

Last synced: 6 days ago
JSON representation

A TensorFlow implementation of Neural Radiance Fields (NeRF) for synthesizing novel views of complex 3D scenes from a sparse set of 2D images.

• Host: GitHub
• URL: https://github.com/alessiopittiglio/dl-project
• Owner: alessiopittiglio
• Created: 2025-06-06T11:15:17.000Z (about 1 year ago)
• Default Branch: main
• Last Pushed: 2025-11-25T02:40:21.000Z (7 months ago)
• Last Synced: 2026-07-01T03:34:30.725Z (6 days ago)
• Topics: computer-vision, nerf, novel-view-synthesis, tensorflow
• Language: Jupyter Notebook
• Homepage:
• Size: 6.38 MB
• Stars: 0
• Watchers: 0
• Forks: 0
• Open Issues: 0
• Metadata Files:
  • Readme: README.md

Awesome Lists containing this project

README

          
# Novel View Synthesis with NeRF

## Project Overview

The goal of this project is to develop a model for Novel View Synthesis. This involves generating new images of a 3D object from viewpoints that are not present in the initial set of images. The model is trained on RGB images and their corresponding camera poses.

## Dataset

The project uses the NeRF Synthetic Dataset provided by Cornell University, specifically the Lego scene.

Dataset structure:

```

nerf_synthetic/

└── SCENE_NAME/

    ├── train/

    ├── val/

    ├── test/

    ├── transforms_train.json

    ├── transforms_val.json

    └── transforms_test.json

```

## Technical Approach

The solution implements the NeRF paper from scratch. The model is Multi-Layer Perceptron (MLP). It takes the encoded position and viewing direction as input and predicts the RGB color and volume density ($\sigma$) at that point. The architecture is composed of 8 fully connected layers with a skip connection at the 4th layer.

## Training

A two-phase strategy was adopted: an initial phase of 50,000 iterations followed by a fine-tuning phase of 10,000 iterations. The model achieved a final PSNR of 30.3 on the test set.

## 360° View Synthesis

A 360° camera rotation generated by the trained NeRF model is shown below.

![NeRF Rotation](assets/rotation.gif)