Context and Geometry Aware Voxel Transformer for Semantic Scene Completion

TL;DR

A novel context and geometry aware voxel transformer that extends deformable cross-attention from 2D to 3D pixel space, enabling the differentiation of points with similar image coordinates based on their depth coordinates and outperforms approaches employing temporal images as inputs or much larger image backbone networks.

Abstract

Vision-based Semantic Scene Completion (SSC) has gained much attention due to its widespread applications in various 3D perception tasks. Existing sparse-to-dense approaches typically employ shared context-independent queries across various input images, which fails to capture distinctions among them as the focal regions of different inputs vary and may result in undirected feature aggregation of cross-attention. Additionally, the absence of depth information may lead to points projected onto the image plane sharing the same 2D position or similar sampling points in the feature map, resulting in depth ambiguity. In this paper, we present a novel context and geometry aware voxel transformer. It utilizes a context aware query generator to initialize context-dependent queries tailored to individual input images, effectively capturing their unique characteristics and aggregating information within the region of interest. Furthermore, it extend deformable cross-attention from 2D to 3D pixel space, enabling the differentiation of points with similar image coordinates based on their depth coordinates. Building upon this module, we introduce a neural network named CGFormer to achieve semantic scene completion. Simultaneously, CGFormer leverages multiple 3D representations (i.e., voxel and TPV) to boost the semantic and geometric representation abilities of the transformed 3D volume from both local and global perspectives. Experimental results demonstrate that CGFormer achieves state-of-the-art performance on the SemanticKITTI and SSCBench-KITTI-360 benchmarks, attaining a mIoU of 16.87 and 20.05, as well as an IoU of 45.99 and 48.07, respectively. Remarkably, CGFormer even outperforms approaches employing temporal images as inputs or much larger image backbone networks.

Figures (7)

• Figure 1: Comparison of feature aggregation. (a) VoxFormer VoxFormer employs a set of shared context-independent queries for different input images, which fails to capture distinctions among them and may lead to undirected feature aggregation. Besides, due to the ignorance of depth information, multiple 3D points may be projected to the same 2D point, causing depth ambiguity. (b) Our CGFormer initializes the voxel queries based on individual input images, effectively capturing their unique features and aggregating information within the region of interest. Furthermore, the deformable cross-attention is extended from 2D to 3D pixel space, enabling the points with similar image coordinates to be distinguished based on their depth coordinates.
• Figure 2: Schematics and detailed architectures of CGFormer. (a) The framework of the proposed CGFormer for camera-based semantic scene completion. The pipeline consists of the image encoder for extracting 2D features, the context and geometry aware voxel (CGVT) transformer for lifting the 2D features to 3D volumes, the 3D local and global encoder (LGE) for enhancing the 3D volumes and a decoding head to predict the semantic occupancy. (b) Detailed structure of the context and geometry aware voxel transformer. (c) Details of the Depth Net.
• Figure 3: Visualization of the sampling locations for different small objects. The yellow dot represents the query point, while the red dots indicate the locations of the deformable sampling points. The sampling points of the context-dependent query (a) tend to be distributed within the regions of interest. Beneficial from this, CGFormer achieve better performance than previous methods.
• Figure 4: Qualitative visualization results on the SemanticKITTI SemanticKITTI validation set.
• Figure 5: Failure cases.