Artificial intelligence

This AI article introduces PaletteNeRF, a new method for photorealistic appearance editing of Neural Radiation Fields (NeRF) based on 3D color decomposition


The ability of Neural Radiance Fields (NeRF) and its derivatives to accurately recreate real-world 3D scenes from 2D photos and to enable new synthesis of high-quality photorealistic views has generated increasing interest in these last years. However, because the appearance of the scene is implicitly registered in neural features and network weights that do not allow local manipulation or intuitive alteration, such volumetric representations are difficult to modify. Several methods have supported NeRF editing. A group of techniques retrieves the material qualities of the scene so that they can be modified, such as surface roughness, or rendered under new lighting conditions.

These techniques depend on an accurate assessment of the scene’s reflectance, which is often difficult for complicated real-world images taken in an open environment. Another class of methods involves discovering latent code that NeRF can be trained to use to achieve the desired appearance. However, these techniques do not offer fine editing and often have limited capacity and flexibility. Also, while some other methods can adapt the appearance of NeRF to suit a certain type of image, they sometimes fail to preserve the same amount of photorealism in the original scene. They suggest PaletteNeRF in this work as an innovative way to facilitate flexible and simple editing of NeRF.

Figure 1: Innovative PaletteNeRFan technique for effective modification of the appearance of the neural radiation field (NeRF). The method reconstructs a NeRF and splits its appearance into a series of (b) color bases based on a 3D palette using (a) multi-view images as training input. This allows to (c) recolor the scene naturally and photos realistically with 3D consistency from any angle. Additionally, they demonstrate that (d) the approach enables several palette-based editing applications, including lighting adjustment and 3D photorealistic style transfer.

Their approach is influenced by earlier image editing techniques that used color palettes, which use a condensed selection of hues to represent the full spectrum of shades in an image. They combine specular and diffuse components to describe the brightness of each point, and they further divide the diffuse component into a linear combination of view-independent common color bases. To reduce the disparity between the produced images and the ground truth images, they jointly optimize the specular component per point, the global color bases and the linear weights per point during training.

To promote scarcity and spatial consistency of decomposition and create more meaningful grouping, they also apply unique regularizers on the weights. By freely modifying the taught color bases, students can intuitively adjust the appearance of NeRF with the suggested framework (Fig. 1). Additionally, they demonstrate how their system can be used in conjunction with semantic features to provide semantic editing. Their technique provides more globally and 3D consistent recoloring outputs of the scene across arbitrary viewpoints than previous palette-based image or video editing techniques. They show that their approach numerically and subjectively outperforms basic approaches, allowing more precise local color modification while faithfully maintaining the photorealism of the 3D scene.

In summary,

• They provide a unique framework for easy modification of NeRF by decomposing the glow field into a weighted mixture of learned color bases.

• To produce intuitive decompositions, they devised a reliable optimization technique using unique regularizers.

• Their method allows for realistic palette-based customization of appearance, allowing even inexperienced users to interact with NeRF in a simple and manageable way on common hardware.

Check Paper and Project. All credit for this research goes to the researchers on this project. Also don’t forget to register. our Reddit page and discord channelwhere we share the latest AI research news, cool AI projects, and more.

Aneesh Tickoo is an intern consultant at MarktechPost. He is currently pursuing his undergraduate studies in Data Science and Artificial Intelligence at Indian Institute of Technology (IIT), Bhilai. He spends most of his time working on projects aimed at harnessing the power of machine learning. His research interest is image processing and is passionate about building solutions around it. He enjoys connecting with people and collaborating on interesting projects.

Leave a Reply