Volume Transformer: Revisiting Vanilla Transformers for 3D Scene Understanding
(ECCV 2026 Long Oral)

This repository contains the official implementation of Volume Transformer (Volt).

Volt partitions the input 3D scene into non-overlapping volumetric patches and embeds each patch into a token with a linear tokenizer. The resulting token sequence is processed by a Transformer encoder with global attention. The latent tokens are then upsampled back to the voxel resolution with a single transposed convolution and mapped to semantic predictions by a linear classification head.

The core Volt model implementation can be found in pointcept/models/volt/volt_base.py.

📢 News

2026-08-05: 🎉🎉 Volt is selected as one of the 28 Long Oral presentations at ECCV 2026!
2026-06-18: 🎉 Volt is accepted to ECCV 2026!
2026-06-13: Volt is now implemented in pure PyTorch and no longer depends on spconv.
2026-06-03: Volt won 3 challenges at CVPR 2026.
2026-04-22: Code release.

Setup

This repository is built on top of Pointcept and incorporates components from SGIFormer for instance segmentation. For integrating image features with 3D backbones, please refer to our DITR codebase.

Dependencies

We recommend using uv, a fast Python package and environment manager, to install the environment.

To install uv on macOS and Linux, run:

curl -LsSf https://astral.sh/uv/install.sh | sh

Then set up the environment with:

# Make sure to load CUDA 12.6 beforehand
# This will automatically create a virtual environment (.venv) and install dependencies from pyproject.toml
uv sync
source .venv/bin/activate

Data Preprocessing

Follow the dataset setup instructions in the Pointcept README.

Indoor Datasets

Preprocessing for ScanNet and ScanNet200 is identical to Pointcept.

ScanNet++ Dataset

After the ECCV submission, we found that the original map_benchmark.csv is not exhaustive, as discussed here. As a result, some labels were mapped to ignore. We therefore regenerate map_benchmark.csv and provide the updated version at pointcept/datasets/preprocessing/scannetpp/new_map_benchmark.csv. Run the preprocessing script below which uses this new mapping.

python pointcept/datasets/preprocessing/scannetpp/preprocess_scannetpp.py --dataset_root ${SCANNETPP_DIR} --output_root ${PROCESSED_SCANNETPP_DIR}

SceneFun3D Dataset

Run the preprocessing script below.

python pointcept/datasets/preprocessing/scenefun3d/preprocess_scenefun3d.py --dataset_root ${SCENEFUN3D_DIR} --output_root ${PROCESSED_SCENEFUN3D_DIR}

Nuscenes

For nuScenes, run the preprocessing script below. Unlike Pointcept preprocessing, we additionally write panoptic labels to the .pkl files.

uv run --no-project --python 3.12 --with nuscenes-devkit python pointcept/datasets/preprocessing/nuscenes/preprocess_nuscenes_info.py --dataset_root ${NUSCENES_DIR} --output_root ${PROCESSED_NUSCENES_DIR}

SemanticKITTI

For SemanticKITTI, run the following script to generate the instance database used for instance CutMix.

python pointcept/datasets/preprocessing/semantic_kitti/build_instance_db_h5.py --dataset_root ${KITTI_DIR} --output_root "data/semantic_kitti_instances"

Waymo

For Waymo, run the preprocessing script below. Waymo provides multiple LiDAR sensors. Unlike Pointcept preprocessing, we use only the points from the TOP LiDAR sensor, since only those points have semantic labels.

uv run --no-project --python 3.10 --with waymo-open-dataset-tf-2-11-0 python pointcept/datasets/preprocessing/waymo/preprocess_waymo.py --dataset_root ${WAYMO_DIR} --output_root ${PROCESSED_WAYMO_DIR} --splits training validation --num_workers ${NUM_WORKERS}

Train

Download UNet teacher weights from HuggingFace

hf download KadirYilmaz/Volt --include "teacher_weights/*.pth" --local-dir weights/

Then, run the training script with the semseg-volt-distill config for each dataset.

### ScanNet
sh scripts/train.sh -g 4 -d scannet -c semseg-volt-distill -n semseg-volt-distill
### ScanNet200
sh scripts/train.sh -g 4 -d scannet200 -c semseg-volt-distill -n semseg-volt-distill
### ScanNet++
sh scripts/train.sh -g 4 -d scannetpp -c semseg-volt-distill -n semseg-volt-distill
### NuScenes
sh scripts/train.sh -g 4 -d nuscenes -c semseg-volt-distill -n semseg-volt-distill
### SemanticKITTI
sh scripts/train.sh -g 4 -d semantic_kitti -c semseg-volt-distill -n semseg-volt-distill
### Waymo
sh scripts/train.sh -g 4 -d waymo -c semseg-volt-distill -n semseg-volt-distill

For joint training, use the semseg-volt-joint-small config instead.

### ScanNet
sh scripts/train.sh -g 4 -d scannet -c semseg-volt-joint-small -n semseg-volt-joint-small
### ScanNet200
sh scripts/train.sh -g 4 -d scannet200 -c semseg-volt-joint-small -n semseg-volt-joint-small
### NuScenes
sh scripts/train.sh -g 4 -d nuscenes -c semseg-volt-joint-small -n semseg-volt-joint-small
### SemanticKITTI
sh scripts/train.sh -g 4 -d semantic_kitti -c semseg-volt-joint-small -n semseg-volt-joint-small
### Waymo
sh scripts/train.sh -g 4 -d waymo -c semseg-volt-joint-small -n semseg-volt-joint-small

Instance Segmentation

First, run the preprocessing script to generate superpoints for ScanNet and ScanNet200.

python pointcept/datasets/preprocessing/scannet/preprocess_superpoints.py --dataset_root ${RAW_SCANNET_DIR} --output_root ${PROCESSED_SCANNET_DIR}

Download the pretrained Volt backbone weights from HuggingFace

mkdir -p weights
curl -L -o weights/volt-small-scannet.pth https://huggingface.co/KadirYilmaz/Volt/resolve/main/Volt_experiments/joint_training_small/scannet/model/model_last.pth
curl -L -o weights/volt-base-scannet.pth https://huggingface.co/KadirYilmaz/Volt/resolve/main/Volt_experiments/joint_training_base/scannet/model/model_last.pth
curl -L -o weights/volt-small-scannet200.pth https://huggingface.co/KadirYilmaz/Volt/resolve/main/Volt_experiments/joint_training_small/scannet200/model/model_last.pth
curl -L -o weights/volt-base-scannet200.pth https://huggingface.co/KadirYilmaz/Volt/resolve/main/Volt_experiments/joint_training_base/scannet200/model/model_last.pth
curl -L -o weights/volt-base-scannetpp.pth https://huggingface.co/KadirYilmaz/Volt/resolve/main/Volt_experiments/joint_training_base/scannetpp/model/model_last.pth

Alternatively you can train them yourself using the corresponding configs above.

Then, run the training script with the insseg-spformer-volt-S-0-base config for scannet/scannet200

### ScanNet
sh scripts/train.sh -g 4 -d scannet -c insseg-spformer-volt-S-0-base -n insseg-volt
### ScanNet200
sh scripts/train.sh -g 4 -d scannet200 -c insseg-spformer-volt-S-0-base -n insseg-volt
### ScanNet++
sh scripts/train.sh -g 4 -d scannetpp -c insseg-spformer-volt-B-0-base -n insseg-volt

For SceneFun3D, we train the model for semantic segmentation only, and use a simple clustering-based post-processing algorithm for instance segmentation. This is done automatically in the SceneFun3DTester and the results are saved in the correct format. We note that we achieve much higher scores on the validation set compared to the test set but we don't have a reason why and we don't know if it is specific to our approach. It could be due to the fact that the test set containing more challenging scenes.

### SceneFun3D
sh scripts/train.sh -g 4 -d scenefun3d -c semseg-volt-base -n insseg-volt

Model Zoo

We provide the experiment directories, including configs, logs, and checkpoints. The experiments can also be seen from Hugging Face.

Semantic Segmentation: Single-Dataset Training

Model	Dataset	Val mIoU	Exp. Dir
Volt-S	ScanNet	77.3	link
Volt-S	ScanNet200	36.1	link
Volt-S	ScanNet++	51.1	link
Volt-S	nuScenes	81.1	link
Volt-S	SemanticKITTI	70.3	link
Volt-S	Waymo	71.2	link

Semantic Segmentation: Joint Training

Model	Dataset	Val mIoU	Exp. Dir
Volt-S	ScanNet	80.2	link
Volt-B	ScanNet	80.5	link
Volt-S	ScanNet200	38.5	link
Volt-B	ScanNet200	40.0	link
Volt-B	ScanNet++	51.7	link
Volt-S	nuScenes	81.8	link
Volt-B	nuScenes	82.2	link
Volt-S	SemanticKITTI	72.8	link
Volt-S	Waymo	72.5	link
Volt-B	Waymo	73.7	link

Instance Segmentation

Model	Dataset	Val mAP50	Exp. Dir
Volt-S	ScanNet	78.5	link
Volt-S	ScanNet200	49.0	link
Volt-B	ScanNet++	53.1	link
Volt-B	SceneFun3D	53.7	link

Citation

If you use our work in your research, please use the following BibTeX entry.

@inproceedings{yilmaz2026volt,
  title     = {{Volume Transformer: Revisiting Vanilla Transformers for 3D Scene Understanding}},
  author    = {Yilmaz, Kadir and Kruse, Adrian and H{\"o}fer, Tristan and de Geus, Daan and Leibe, Bastian},
  booktitle = {European Conference on Computer Vision (ECCV)},
  year      = {2026},
}

Name		Name	Last commit message	Last commit date
.github/workflows		.github/workflows
configs		configs
data		data
libs		libs
pointcept		pointcept
scripts		scripts
tools		tools
weights		weights
.gitignore		.gitignore
LICENSE		LICENSE
README.md		README.md
pyproject.toml		pyproject.toml
uv.lock		uv.lock

Navigation Menu

Repository navigation

About

Repository files navigation

Volume Transformer: Revisiting Vanilla Transformers for 3D Scene Understanding
(ECCV 2026 Long Oral)

📢 News

Setup

Dependencies

Data Preprocessing

Indoor Datasets

ScanNet++ Dataset

SceneFun3D Dataset

Nuscenes

SemanticKITTI

Waymo

Train

Instance Segmentation

Model Zoo

Semantic Segmentation: Single-Dataset Training

Semantic Segmentation: Joint Training

Instance Segmentation

Citation

About

Releases

Packages

Contributors

Languages

Latest commit

History

Folders and files

Repository files navigation

Volume Transformer: Revisiting Vanilla Transformers for 3D Scene Understanding (ECCV 2026 Long Oral)

📢 News

Setup

Dependencies

Data Preprocessing

Indoor Datasets

ScanNet++ Dataset

SceneFun3D Dataset

Nuscenes

SemanticKITTI

Waymo

Train

Instance Segmentation

Model Zoo

Semantic Segmentation: Single-Dataset Training

Semantic Segmentation: Joint Training

Instance Segmentation

Citation

About

Topics

Resources

Stars

Watchers

Forks

Releases

Packages

Contributors

Languages

Volume Transformer: Revisiting Vanilla Transformers for 3D Scene Understanding
(ECCV 2026 Long Oral)