PolarFormer: Multi-camera 3D Object Detection with Polar Transformer

TL;DR

PolarFormer tackles 3D object detection from multi-camera images by shifting from Cartesian to Polar coordinates, aligning with the ego car’s wedge-shaped imaging geometry. It introduces a cross-attention based cross-plane encoder to map image columns to Polar rays, a Polar alignment module to fuse data across cameras, and a multi-scale Polar BEV encoder/decoder to handle object scale variations along the radial axis. The approach yields a structured Polar BEV representation that is decoded directly in Polar coordinates, eliminating depth estimation and post-processing steps. On nuScenes, PolarFormer achieves state-of-the-art mAP and NDS, with PolarFormer-T further improving performance through temporal fusion, demonstrating the practicality of Polar-centric BEV reasoning for camera-based 3D detection.

Abstract

3D object detection in autonomous driving aims to reason "what" and "where" the objects of interest present in a 3D world. Following the conventional wisdom of previous 2D object detection, existing methods often adopt the canonical Cartesian coordinate system with perpendicular axis. However, we conjugate that this does not fit the nature of the ego car's perspective, as each onboard camera perceives the world in shape of wedge intrinsic to the imaging geometry with radical (non-perpendicular) axis. Hence, in this paper we advocate the exploitation of the Polar coordinate system and propose a new Polar Transformer (PolarFormer) for more accurate 3D object detection in the bird's-eye-view (BEV) taking as input only multi-camera 2D images. Specifically, we design a cross attention based Polar detection head without restriction to the shape of input structure to deal with irregular Polar grids. For tackling the unconstrained object scale variations along Polar's distance dimension, we further introduce a multi-scalePolar representation learning strategy. As a result, our model can make best use of the Polar representation rasterized via attending to the corresponding image observation in a sequence-to-sequence fashion subject to the geometric constraints. Thorough experiments on the nuScenes dataset demonstrate that our PolarFormer outperforms significantly state-of-the-art 3D object detection alternatives.

Figures (10)

• Figure 1: Taking multi-camera images as input, the proposed PolarFormer model is designed particularly for accurate 3D object detection in the Polar coordinate system.
• Figure 2: Schematic illustration of our proposed PolarFormer for multi-camera 3D object detection. For each image captured by any camera view, our model first extracts the feature maps at multiple spatial scales. Given such a feature map, the cross-plane encoder (a) then transforms all the feature columns to a set of Polar rays in a sequence-to-sequence manner via polar queries based cross-attention. The polar rays from all the cameras are subsequently processed by a Polar alignment module (b) to generate a structured multi-scale Polar BEV map, followed by further enhancement via interactions among different scales using a Polar BEV encoder (c). At last, a specially designed Polar Head decodes multi-scale Polar BEV features for making final predictions in the Polar coordinate.
• Figure 3: Cartesian and Polar coordinates.
• Figure 4: Multi-scale Polar BEV maps.
• Figure 5: 3D object detection in (a) Cartesian BEV vs. (b) Polar BEV, and (c) Performance comparison (mAP/NDS) at three distances (Near/Medium/Far). Red and green boxes show the same objects in different coordinates.