# Volt

A research implementation of a transformer architecture for 3D scene understanding, partitioning a scene into volumetric patches for semantic and instance segmentation.

Volt, short for Volume Transformer, is the official code release for an ECCV 2026 paper on 3D scene understanding. It divides an input 3D scene into non-overlapping volumetric patches, embeds each patch into a token with a linear tokenizer, and processes the resulting sequence with a standard transformer encoder using global attention.

The latent tokens are upsampled back to voxel resolution with a single transposed convolution and mapped to semantic predictions by a linear classification head. The implementation builds on the Pointcept framework and incorporates components from SGIFormer for instance segmentation, with additional support for combining image features with 3D backbones. It targets datasets including ScanNet, ScanNet200, ScanNet++, and SceneFun3D, and is built with PyTorch, installed and run through the uv package manager.

Website: https://www.vision.rwth-aachen.de/Volt