AI has revolutionized the world of 3D modeling with its latest development: Neuralangelo: High-Fidelity Neural Surface Reconstruction by Zhaoshuo Li and colleagues from NVIDIA Research and Johns Hopkins University. This paper presents a novel framework for creating detailed 3D models of real-world scenes from RGB images using neural volume rendering. Let me explain why this is so cool and how it works.

Photogrammetry, for those unfamiliar with the term, is the combination of art and science that involves measuring objects in the real world using images and other sensors like LiDAR. NVIDIA’s Neuralangelo essentially amplifies the capabilities of photogrammetry by leveraging the power of AI.

Traditional photogrammetry has limitations when it comes to handling repetitive structures, textureless surfaces, and strong color variations. However, Neuralangelo overcomes these obstacles by incorporating the technology behind Instant NeRF, enabling it to capture even the most intricate details.

Until now, NeRFs (Neural Radiance Fields) and photogrammetry have served different purposes. NeRFs are renowned for producing stunning visualizations, such as captivating flythroughs. However, when transformed into 3D meshes, they lack surface detail. On the other hand, photogrammetry excels at surface reconstruction and precise measurements, but its visual appeal can sometimes be lacking.

Neuralangelo emerges as the ultimate game-changer. Its neural approach bridges the gap between visual aesthetics and surface reconstruction, providing the best of both worlds. Gone are the days of blurry or incomplete results—Neuralangelo delivers crisp and highly detailed 3D surfaces that are visually enticing and leave a lasting impression.

How it works:

Neural volume rendering is a technique that uses a neural network to learn a function that maps 3D coordinates to colors and densities. This function can then be used to render novel views of the scene from any camera pose. Unlike traditional photogrammetry methods that rely on multi-view stereo (MVS) algorithms to estimate depth maps and reconstruct surfaces, neural volume rendering does not require any explicit geometry representation or auxiliary data such as segmentation or depth. This makes it more robust to challenging scenes with occlusions, textureless regions, or complex lighting.

However, existing neural volume rendering methods have some limitations. First, they often use low-resolution voxel grids or octrees to store the learned function, which limits the resolution and detail of the reconstructed scene. Second, they use gradient descent to optimize the function, which can be slow and prone to local minima. Third, they do not explicitly model the surface of the scene, which can result in blurry or distorted renderings.

Neuralangelo addresses these issues by combining two key ideas: multi-resolution 3D hash grids and numerical gradients. A 3D hash grid is a data structure that allows efficient storage and retrieval of sparse 3D data. Neuralangelo uses multiple hash grids with different resolutions to represent the scene at different levels of detail. This allows it to capture fine-grained structures without wasting memory on empty space. A numerical gradient is a way of approximating the derivative of a function using finite differences. Neuralangelo uses numerical gradients to compute higher-order derivatives of the learned function, such as curvature and normal vectors. These derivatives are used as a smoothing operation to regularize the optimization and improve the quality of the surface.

Neuralangelo also uses a coarse-to-fine optimization strategy on the hash grids. It starts with a low-resolution grid and gradually increases the resolution as the optimization progresses. This helps it avoid local minima and converge faster. The final result is a high-fidelity neural surface reconstruction that can render realistic and detailed views of the scene from any angle.

To demonstrate the effectiveness of Neuralangelo, the authors applied it to various real-world scenes captured by RGB cameras, such as buildings, statues, cars, and people. They compared it with state-of-the-art methods for neural volume rendering and MVS-based surface reconstruction. They showed that Neuralangelo can produce more accurate and detailed reconstructions than previous methods, even without any auxiliary data such as depth or segmentation. They also showed that Neuralangelo can handle large-scale scenes with complex geometry and lighting, such as a courthouse or a cathedral.

Neuralangelo is a breakthrough in neural surface reconstruction that opens up new possibilities for 3D content creation and digital twin applications. It enables anyone with a smartphone or a camera to capture and reconstruct high-quality 3D models of their surroundings with minimal effort. It also provides a powerful tool for researchers and artists who want to explore and manipulate 3D scenes using neural networks.


In summary, Neuralangelo represents a significant leap forward in achieving realistic visuals combined with finely detailed 3D models that accurately represent surfaces. This breakthrough opens up a world of possibilities for applications such as digital twins, gaming, and visual effects. NVIDIA has once again demonstrated their innovative prowess with this remarkable technology.

If you are interested in learning more about Neuralangelo, you can check out the paper or the project page for more details and videos. You can also try it out yourself by downloading the code and running it on your own images. I hope you enjoyed this blog post and found it informative and inspiring. Stay tuned for more updates on 3D graphics and computer vision research!