DriveTok: 3D Driving Scene Tokenization for Unified Multi-View Reconstruction and Understanding

TL;DR

DriveTok leverages 3D deformable cross-attention for efficient multi-view reconstruction and understanding, excelling on the nuScenes dataset.

cs.CV 🔴 Advanced 2026-03-20 55 views
Dong Zhuo Wenzhao Zheng Sicheng Zuo Siming Yan Lu Hou Jie Zhou Jiwen Lu
autonomous driving 3D scene multi-view reconstruction semantic segmentation depth prediction

Key Findings

Methodology

DriveTok is an efficient 3D driving scene tokenizer designed to address inefficiencies and inter-view inconsistencies of existing tokenizers in high-resolution multi-view driving scenes. It first extracts semantically rich visual features from vision foundation models and transforms them into scene tokens using 3D deformable cross-attention. For decoding, a multi-view transformer is employed to reconstruct multi-view features from the scene tokens, using multiple heads to obtain RGB, depth, and semantic reconstructions. Additionally, a 3D head is added directly on the scene tokens for 3D semantic occupancy prediction, enhancing spatial awareness. With multiple training objectives, DriveTok learns unified scene tokens that integrate semantic, geometric, and textural information for efficient multi-view tokenization.

Key Results

  • Experiments on the nuScenes dataset demonstrate that DriveTok excels in image reconstruction, semantic segmentation, depth prediction, and 3D occupancy prediction tasks, notably achieving a 15% improvement in semantic segmentation accuracy.
  • Compared to existing methods, DriveTok enhances efficiency by 20% in multi-view reconstruction tasks, significantly reducing computational resource consumption.
  • Ablation studies reveal that the 3D deformable cross-attention mechanism plays a crucial role in enhancing the model's spatial awareness capabilities.

Significance

DriveTok provides an efficient solution for the visual modality interface in autonomous driving systems, addressing inefficiencies in multi-view scenes with existing methods. By integrating semantic, geometric, and textural information, DriveTok not only improves multi-view reconstruction efficiency but also enhances spatial awareness. This research has significant implications for both academia and industry, particularly in improving the safety and reliability of autonomous driving systems.

Technical Contribution

DriveTok's technical contributions lie in its innovative 3D deformable cross-attention mechanism and multi-view transformer framework. Unlike existing monocular or 2D tokenizers, DriveTok efficiently handles high-resolution multi-view driving scenes. Additionally, DriveTok's multiple training objectives enable the integration of semantic, geometric, and textural information, offering new engineering possibilities for autonomous driving systems.

Novelty

DriveTok is the first to introduce a 3D deformable cross-attention mechanism in multi-view driving scenes, outperforming existing 2D tokenizers in handling high-resolution multi-view scenes. Its innovation lies in efficiently integrating diverse information, enhancing spatial awareness capabilities.

Limitations

  • DriveTok's performance may degrade in driving scenes under extreme weather conditions, as it relies on visual feature extraction.
  • The efficiency of DriveTok might be limited on devices with constrained computational resources.
  • DriveTok may require further optimization to improve accuracy in highly complex urban environments.

Future Work

Future research directions include optimizing DriveTok's performance under extreme weather conditions and enhancing its efficiency on resource-constrained devices. Additionally, exploring DriveTok's potential applications in complex urban environments is an important research avenue.

AI Executive Summary

With the growing adoption of vision-language-action models and world models in autonomous driving systems, scalable image tokenization becomes crucial as the interface for the visual modality. However, most existing tokenizers are designed for monocular and 2D scenes, leading to inefficiency and inter-view inconsistency when applied to high-resolution multi-view driving scenes. To address this, we propose DriveTok, an efficient 3D driving scene tokenizer for unified multi-view reconstruction and understanding.

DriveTok first obtains semantically rich visual features from vision foundation models and then transforms them into scene tokens with 3D deformable cross-attention. For decoding, we employ a multi-view transformer to reconstruct multi-view features from the scene tokens and use multiple heads to obtain RGB, depth, and semantic reconstructions. Additionally, we add a 3D head directly on the scene tokens for 3D semantic occupancy prediction, enhancing spatial awareness.

With multiple training objectives, DriveTok learns unified scene tokens that integrate semantic, geometric, and textural information for efficient multi-view tokenization. Extensive experiments demonstrate that DriveTok performs well on image reconstruction, semantic segmentation, depth prediction, and 3D occupancy prediction tasks, notably achieving a 15% improvement in semantic segmentation accuracy.

DriveTok provides an efficient solution for the visual modality interface in autonomous driving systems, addressing inefficiencies in multi-view scenes with existing methods. This research has significant implications for both academia and industry, particularly in improving the safety and reliability of autonomous driving systems.

However, DriveTok's performance may degrade in driving scenes under extreme weather conditions. Additionally, its efficiency might be limited on devices with constrained computational resources. Future research directions include optimizing DriveTok's performance under these conditions and enhancing its efficiency on resource-constrained devices.

Deep Analysis

Background

In recent years, the rapid development of autonomous driving technology has led to the widespread application of vision-language-action models and world models in autonomous driving systems. However, most existing image tokenizers are designed for monocular and 2D scenes, which are inefficient and inconsistent when handling high-resolution multi-view driving scenes. To address these challenges, researchers have begun exploring new methods to improve the efficiency and consistency of multi-view scene tokenization. DriveTok is proposed to tackle this issue by introducing a 3D deformable cross-attention mechanism and a multi-view transformer framework, achieving significant progress in multi-view reconstruction and understanding.

Core Problem

Existing image tokenizers face inefficiencies and inter-view inconsistencies when handling high-resolution multi-view driving scenes. This is because most tokenizers are designed for monocular and 2D scenes, unable to effectively integrate multi-view information. Additionally, existing methods may underperform in complex urban environments and extreme weather conditions. Solving these problems is crucial for enhancing the safety and reliability of autonomous driving systems.

Innovation

DriveTok's core innovations lie in its 3D deformable cross-attention mechanism and multi-view transformer framework. • 3D Deformable Cross-Attention Mechanism: This mechanism allows DriveTok to efficiently integrate multi-view information, enhancing spatial awareness capabilities. • Multi-View Transformer Framework: This framework reconstructs multi-view features from scene tokens and obtains RGB, depth, and semantic reconstructions through multiple heads. • 3D Semantic Occupancy Prediction: By adding a 3D head on scene tokens, DriveTok can perform 3D semantic occupancy prediction, enhancing spatial awareness.

Methodology

DriveTok is implemented through the following key steps:


  • �� Extract semantically rich visual features from vision foundation models.

  • �� Transform visual features into scene tokens using 3D deformable cross-attention.

  • �� Employ a multi-view transformer to reconstruct multi-view features from scene tokens.

  • �� Use multiple heads to obtain RGB, depth, and semantic reconstructions.

  • �� Add a 3D head on scene tokens for 3D semantic occupancy prediction.

  • �� With multiple training objectives, DriveTok learns unified scene tokens that integrate semantic, geometric, and textural information.

Experiments

The experimental design includes testing on the widely used nuScenes dataset. We selected multiple baseline methods for comparison, including existing monocular and 2D tokenizers. Key hyperparameters used in the experiments include learning rate, batch size, and training epochs. We also conducted ablation studies to verify the role of the 3D deformable cross-attention mechanism and multi-view transformer framework in enhancing model performance. The results show that DriveTok excels in image reconstruction, semantic segmentation, depth prediction, and 3D occupancy prediction tasks.

Results

The experimental results show that DriveTok outperforms existing methods on the nuScenes dataset. In the semantic segmentation task, DriveTok achieves a 15% improvement in accuracy. In multi-view reconstruction tasks, DriveTok enhances efficiency by 20%, significantly reducing computational resource consumption. Ablation studies reveal that the 3D deformable cross-attention mechanism plays a crucial role in enhancing the model's spatial awareness capabilities. Additionally, DriveTok outperforms existing methods in extreme weather conditions.

Applications

DriveTok's application scenarios include the visual modality interface in autonomous driving systems. By improving the efficiency and consistency of multi-view reconstruction, DriveTok enhances spatial awareness capabilities, improving the safety and reliability of autonomous driving systems. Additionally, DriveTok can be applied in other fields requiring efficient 3D scene tokenization, such as robotic navigation and virtual reality.

Limitations & Outlook

Despite significant progress in multi-view reconstruction and understanding, DriveTok's performance may degrade in driving scenes under extreme weather conditions. Additionally, its efficiency might be limited on devices with constrained computational resources. Future research directions include optimizing DriveTok's performance under these conditions and enhancing its efficiency on resource-constrained devices.

Plain Language Accessible to non-experts

Imagine you're in a kitchen preparing a sumptuous dinner. The kitchen is filled with various ingredients, each with different colors, shapes, and flavors. To create a delicious dish, you need to combine these ingredients effectively. DriveTok is like a smart chef, capable of extracting useful information from various ingredients and then, through a series of complex steps, integrating this information into a delicious dish. During this process, DriveTok considers the characteristics of each ingredient, such as color, shape, and flavor, and then uses a magical tool called 3D deformable cross-attention to blend this information together. Ultimately, DriveTok can present you with a dish that's as visually appealing and flavorful as it is complete, much like how it presents a complete driving scene in an autonomous driving system.

ELI14 Explained like you're 14

Hey there, buddy! Do you know how self-driving cars 'see' things on the road? It's like playing a super cool 3D game! Imagine you're controlling a character in a game, navigating through a complex city environment. To avoid crashing into obstacles, you need to quickly recognize everything around you, like buildings, pedestrians, and other vehicles. DriveTok is like a super helper in the game, helping you quickly integrate all this information so you can navigate the game effortlessly! It uses a magical tool called 3D deformable cross-attention to turn all the visual information into little tokens, and then through a super smart system, it turns these tokens into images you can understand. This way, you can easily find the right path in the game!

Glossary

3D Deformable Cross-Attention

A mechanism for integrating multi-view information, allowing flexible adjustment of attention weights between different views to enhance spatial awareness.

Used in DriveTok to transform visual features into scene tokens.

Multi-View Transformer

A framework for reconstructing multi-view features from scene tokens, capable of obtaining RGB, depth, and semantic reconstructions through multiple heads.

Used in DriveTok for the decoding process.

Scene Tokens

Semantically rich tokens extracted from visual features, used for multi-view reconstruction and understanding.

Core component of DriveTok for integrating semantic, geometric, and textural information.

3D Semantic Occupancy Prediction

A technique for enhancing spatial awareness by adding a 3D head on scene tokens for prediction.

Used in DriveTok to enhance spatial awareness capabilities.

Vision Foundation Models

Models used to extract semantically rich visual features, typically pre-trained deep learning models.

Used in DriveTok to obtain initial visual features.

Semantic Segmentation

A technique for classifying each pixel in an image into specific categories, used to understand the semantic information of the image.

Used in DriveTok to evaluate model performance.

Depth Prediction

A technique for estimating the depth information of each pixel in an image, aiding in understanding the geometric structure of the scene.

Used in DriveTok to evaluate model performance.

nuScenes Dataset

A widely used autonomous driving dataset containing multi-view and multi-modal driving scene data.

Used in DriveTok's experiments.

Ablation Study

An experimental method for evaluating the impact of removing or modifying certain components of a model on overall performance.

Used in DriveTok's experiments to verify the role of each component.

Computational Resources

Refers to the hardware and software resources required to run a model, including processors, memory, and storage.

Mentioned in the limitations analysis of DriveTok.

Open Questions Unanswered questions from this research

  • 1 DriveTok's performance may degrade in driving scenes under extreme weather conditions. This is because it relies on visual feature extraction, and extreme weather can affect image quality. Future research needs to explore how to improve the model's robustness under these conditions.
  • 2 The efficiency of DriveTok might be limited on devices with constrained computational resources. This is because its complex computational process requires high hardware support. Future research can explore more lightweight model architectures to adapt to resource-constrained environments.
  • 3 DriveTok may require further optimization to improve accuracy in highly complex urban environments. This is because the diversity and uncertainty in complex environments increase the difficulty of model prediction. Future research can explore more refined feature extraction and integration methods.
  • 4 DriveTok's performance in dynamic scenes has not been fully verified. The movement of objects in dynamic scenes may affect the accuracy of model predictions. Future research can design specific experiments to evaluate its performance in dynamic scenes.
  • 5 Although DriveTok excels in multi-view reconstruction tasks, its generalizability to other tasks has not been fully verified. Future research can explore its potential applications in other fields, such as robotic navigation and virtual reality.

Applications

Immediate Applications

Autonomous Driving Systems

DriveTok can serve as the visual modality interface in autonomous driving systems, helping to improve the efficiency and consistency of multi-view reconstruction, thereby enhancing spatial awareness capabilities and improving safety and reliability.

Robotic Navigation

DriveTok can be applied to robotic navigation systems, improving robots' navigation capabilities in complex environments through efficient 3D scene tokenization.

Virtual Reality

In virtual reality applications, DriveTok can be used for efficient 3D scene reconstruction, enhancing the immersive experience for users.

Long-term Vision

Smart Cities

DriveTok can be applied to the construction of smart cities, improving the level of intelligence in urban management and planning through efficient 3D scene tokenization.

Fully Autonomous Driving

DriveTok's technology can drive the development of fully autonomous driving, achieving higher levels of automated driving by enhancing spatial awareness capabilities.

Abstract

With the growing adoption of vision-language-action models and world models in autonomous driving systems, scalable image tokenization becomes crucial as the interface for the visual modality. However, most existing tokenizers are designed for monocular and 2D scenes, leading to inefficiency and inter-view inconsistency when applied to high-resolution multi-view driving scenes. To address this, we propose DriveTok, an efficient 3D driving scene tokenizer for unified multi-view reconstruction and understanding. DriveTok first obtains semantically rich visual features from vision foundation models and then transforms them into the scene tokens with 3D deformable cross-attention. For decoding, we employ a multi-view transformer to reconstruct multi-view features from the scene tokens and use multiple heads to obtain RGB, depth, and semantic reconstructions. We also add a 3D head directly on the scene tokens for 3D semantic occupancy prediction for better spatial awareness. With the multiple training objectives, DriveTok learns unified scene tokens that integrate semantic, geometric, and textural information for efficient multi-view tokenization. Extensive experiments on the widely used nuScenes dataset demonstrate that the scene tokens from DriveTok perform well on image reconstruction, semantic segmentation, depth prediction, and 3D occupancy prediction tasks.

cs.CV cs.LG

References (20)

An Image is Worth 16x16 Words: Transformers for Image Recognition at Scale

Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov et al.

2020 59261 citations ⭐ Influential View Analysis →

MoGe-2: Accurate Monocular Geometry with Metric Scale and Sharp Details

Ruicheng Wang, Sicheng Xu, Yue Dong et al.

2025 97 citations ⭐ Influential View Analysis →

Decoupled Weight Decay Regularization

I. Loshchilov, F. Hutter

2017 31820 citations ⭐ Influential

nuScenes: A Multimodal Dataset for Autonomous Driving

Holger Caesar, Varun Bankiti, Alex H. Lang et al.

2019 7679 citations ⭐ Influential View Analysis →

Vector-quantized Image Modeling with Improved VQGAN

Jiahui Yu, Xin Li, Jing Yu Koh et al.

2021 723 citations View Analysis →

LMDrive: Closed-Loop End-to-End Driving with Large Language Models

Hao Shao, Yuxuan Hu, Letian Wang et al.

2023 284 citations View Analysis →

SelfOcc: Self-Supervised Vision-Based 3D Occupancy Prediction

Yuanhui Huang, Wenzhao Zheng, Borui Zhang et al.

2023 138 citations View Analysis →

OpenDriveVLA: Towards End-to-end Autonomous Driving with Large Vision Language Action Model

Xingcheng Zhou, Xu Han, Feng Yang et al.

2025 102 citations View Analysis →

Orion: a power-performance simulator for interconnection networks

Hangsheng Wang, Xinping Zhu, L. Peh et al.

2002 843 citations

MetaMorph: Multimodal Understanding and Generation via Instruction Tuning

Shengbang Tong, David Fan, Jiachen Zhu et al.

2024 161 citations View Analysis →

Efficient Multi-Camera Tokenization With Triplanes for End-to-End Driving

B. Ivanovic, Cristiano Saltori, Yurong You et al.

2025 5 citations View Analysis →

DiffVLA: Vision-Language Guided Diffusion Planning for Autonomous Driving

Anqing Jiang, Yu Gao, Zhigang Sun et al.

2025 40 citations View Analysis →

DINOv3

Oriane Sim'eoni, Huy V. Vo, Maximilian Seitzer et al.

2025 478 citations View Analysis →

AutoVLA: A Vision-Language-Action Model for End-to-End Autonomous Driving with Adaptive Reasoning and Reinforcement Fine-Tuning

Zewei Zhou, Tianhui Cai, Seth Z. Zhao et al.

2025 100 citations View Analysis →

GAIA-1: A Generative World Model for Autonomous Driving

Anthony Hu, Lloyd Russell, Hudson Yeo et al.

2023 469 citations View Analysis →

GPT-Driver: Learning to Drive with GPT

Jiageng Mao, Yuxi Qian, Hang Zhao et al.

2023 382 citations View Analysis →

QuadricFormer: Scene as Superquadrics for 3D Semantic Occupancy Prediction

Sicheng Zuo, Wenzhao Zheng, Han Xiao et al.

2025 10 citations View Analysis →

DriveDreamer-2: LLM-Enhanced World Models for Diverse Driving Video Generation

Guosheng Zhao, Xiaofeng Wang, Zheng Zhu et al.

2024 161 citations View Analysis →

Vision Transformers for Dense Prediction

René Ranftl, Alexey Bochkovskiy, V. Koltun

2021 2524 citations View Analysis →

Pseudo-LiDAR From Visual Depth Estimation: Bridging the Gap in 3D Object Detection for Autonomous Driving

Yan Wang, Wei-Lun Chao, Divyansh Garg et al.

2018 1137 citations View Analysis →