FEW-SHOT SCENE ACQUISITION FOR GAME ENVIRONMENTS CREATION

点击翻译按钮获取中文摘要

This project aims to create an affordable method of obtaining 3D assets for game development. The method will consist of reconstructing a real scene from few images with physically realistic materials for rendering. Just imagine using a few images from the internet or your camera to create an asset for your game! Modern 3d content production plays a crucial role in the games making and VR/AR applications. However, the entry threshold into this field is currently too high. This affects companies and independent developers alike, requiring a lot of work to master complex 3D software packages as well as sophisticated geometry and material capture setups. This project proposes a way to democratize this field for widespread use and make faster creation of game environments possible. The use of learnable priors will allow complex scenes to be reconstructed from a few images to render the scene in-game. Volumetric rendering techniques with implicit representations look like the best starting point here due to the ease of differentiation and photorealism. The final model can be used as a game renderer or an intermediate representation of the received scene, converted into classic meshes and textures at the end.