Unifying Voxel-based Representation with Transformer for 3D Object Detection

TL;DR

This work tackles multi-modality 3D object detection by unifying LiDAR and camera data in a voxel-based representation. It introduces UVTR, which builds modality-specific voxel spaces $V_I$ and $V_P$ and a unified space $V_U$ to enable direct spatial interactions without height compression, supplemented by cross-modality knowledge transfer and modality fusion. Object-level predictions are produced via a deformable transformer decoder that samples features at reference points $p=(x,y,z)$ in $V_U$, with training guided by a set-to-set loss ${\mathcal{L}}_{Det}$ and an auxiliary ${\mathcal{L}}_{KT}$ when transfer is used. The approach achieves leading results on nuScenes for detection and tracking, demonstrates robustness to sensor perturbations, and highlights the practical benefits of a unified voxel-space framework for multi-sensor 3D perception. ${V_I}$, ${V_P}$, and ${V_U}$ are central to enabling explicit cross-modality interactions and efficient object-level reasoning in 3D space.

Abstract

In this work, we present a unified framework for multi-modality 3D object detection, named UVTR. The proposed method aims to unify multi-modality representations in the voxel space for accurate and robust single- or cross-modality 3D detection. To this end, the modality-specific space is first designed to represent different inputs in the voxel feature space. Different from previous work, our approach preserves the voxel space without height compression to alleviate semantic ambiguity and enable spatial connections. To make full use of the inputs from different sensors, the cross-modality interaction is then proposed, including knowledge transfer and modality fusion. In this way, geometry-aware expressions in point clouds and context-rich features in images are well utilized for better performance and robustness. The transformer decoder is applied to efficiently sample features from the unified space with learnable positions, which facilitates object-level interactions. In general, UVTR presents an early attempt to represent different modalities in a unified framework. It surpasses previous work in single- or multi-modality entries. The proposed method achieves leading performance in the nuScenes test set for both object detection and the following object tracking task. Code is made publicly available at https://github.com/dvlab-research/UVTR.

Figures (7)

• Figure 1: Toy example of methods for unified representation. Compared with others, the proposed manner in \ref{['fig:intro_voxel']} constructs the voxel space by sampling features from the image plane and represents multi-modalities uniformly without height-level compression in \ref{['fig:intro_bev']} that brings semantic ambiguity.
• Figure 2: The framework of UVTR with multi-modality input. Given single- or multi-frame images and point clouds, we first process them in individual backbone and convert to modality-specific space $\mathbf{V}_I$ and $\mathbf{V}_P$, where view transform is utilized for that of image. In voxel encoder, features are spatially interacted, and knowledge transfer is easily supported during training. Single- or multi-modality features are selected via modality switch according to different settings. Finally, transformer decoder is utilized for prediction by sampling features from the unified space $\mathbf{V}_U$ with learnable positions.
• Figure 3: Details in the view transform.
• Figure 4: Details in the knowledge transfer.
• Figure 5: We validate the robustness of UVTR by adding two typical errors during inference. For dropped view in \ref{['fig:drop_view']}, we randomly drop a fixed number of camera views to simulate the camera failure. For sensor noise in \ref{['fig:trans_noise']}, we randomly add translational noises in LiDAR to camera calibration matrix.