Experiment Test for PyTorch Version Link to heading

Code Section Link to heading

1. Program Link to heading

Source Code

2. Description Link to heading

  1. nerf_synthetic Folder

    • Concept: A synthetic dataset rendered using Blender software. To train NeRF, 3D models are created, virtual camera angles are set, and images are rendered from multiple perspectives.
    • Purpose: Used for training and validation.

    Folder contents:

    • The test folder contains test datasets: depth maps, pseudo-rgb, and rgb images.
    • The train folder contains training datasets.
    • The val folder contains validation datasets.
    • *.json files contain camera pose information.

  2. nerf_llff_data Folder

    • Concept: A real-world scene dataset, obtained using the LLFF (Light Field Photography) method. This data comes from real cameras capturing scenes. LLFF collects light field information through images captured from various camera angles, each image having corresponding camera parameters.
    • Purpose: Used for validation and testing.

    Folder contents:

    • The images folder contains images of real-world scenes.
    • The sparse folder contains sparse point cloud data of the real-world scenes.
    • pose_bounds.npy contains camera pose information.
    • simplices.npy contains information about the simplicial complex of the scene.

3. Issues Link to heading

None.

4. Local Reproduction Link to heading

  1. lego_synthetic Dataset
    PSNR = 32 | Iterations = 150,000 | LOSS = 0.0012

  2. fern_llff Dataset
    PSNR = 29 | Iterations = 200,000 | LOSS = 0.0032

Paper Section Link to heading

1. Official Website Link to heading

NeRF Official Website

2. Paper Logic Link to heading

+——————————+ +———————+ +————————-+ +———————-+ | Input Images (Multiple Views)| —> | Neural Network Module (MLP) | —> | Volume Rendering (Volume Ray | —> | Generate New View Image | | Camera Pose Information | | Input: Position (x, y, z) | | Marching) | | Output: New View Image | | | | and View Direction (θ, φ) | | Output: Color (c) and Volume Density | | | +—————————-+ | Output: Color (c) and | | (σ) | +———————-+ | Volume Density (σ) | +————————-+ +———————+ | V +—————————–+ | Neural Network Training (Optimizing Network Parameters) | | Input: Images, Camera Poses, Initial Network | | Output: Optimal Network Parameters | +—————————–+