A Study on Efficient BEV Transformation via Image Compression and Performance Enhancement of Transformer-based 3D Object Detection

Abstract

This study proposes a width-aligned transformer (WAFormer) based bird’s eye view (BEV) perception pipeline for 3D object detection using a multi-view camera. The proposed transformation method compresses multi-view image features along the vertical axis into 1D features, considerably reducing transformer computation. A refine transformer with self- and cross-attention restores vertically compressed information while preserving efficiency. To enhance spatial representation, we introduce a depth-altitude positional encoding that integrates depth distribution and ground-referenced altitude cues via height-aware aggregation and bidirectional cross-attention. The overall network consists of a residual neural network (ResNet-101) backbone, the proposed view transformer, a BEV encoder, and a CenterPoint-style detection head, with an auxiliary fully convolutional, one-stage, object-detection based 1D (used only during training, and temporal alignment) using ego-motion for spatiotemporal feature construction. Under an identically trained baseline condition, the proposed method improves mAP, NDS, and inference speed compared with BEVDet4D. We also analyze the trade-off of vertical compression, that improves detection accuracy and velocity estimation but increases scale and orientation errors.