This repository provides the official code release for the "Simple-BEV" paper, focusing on multi-sensor Bird's-Eye View (BEV) perception for autonomous driving. It offers a baseline implementation for researchers and engineers to evaluate and build upon BEV perception techniques, particularly those leveraging camera and radar data.

How It Works

The project implements a BEV perception system that processes multi-sensor inputs (primarily camera and radar) to generate a unified BEV representation. It utilizes a ResNet-101 encoder for feature extraction from camera images and incorporates radar data, including meta-radar information, with a specified number of past sweeps (nsweeps). The architecture is designed for efficiency and interpretability, allowing for a clear understanding of how different sensor modalities contribute to the final BEV output.

Quick Start & Requirements

• Installation: Use Conda to create and activate an environment (conda create --name bev, source activate bev), then install PyTorch, torchvision, cudatoolkit, and other dependencies via pip install -r requirements.txt.
• Data: Requires downloading the nuScenes dataset and its dependencies.
• Pre-trained Models: Download camera-only or camera-plus-radar models using get_rgb_model.sh and get_rad_model.sh.
• Hardware: Requires NVIDIA GPU with CUDA 11.3.
• Documentation: [ Paper ] [ Project Page ]

Highlighted Details

• Achieves 47.6 mIoU (camera-only) and 55.8 mIoU (camera-plus-radar) on nuScenes at res_scale=2 (448x800).
• Supports training and evaluation for both camera-only and camera-plus-radar configurations.
• Defines clear tensor axis ordering (B,S,C,Z,Y,X) and geometry conventions for transformations.
• Includes sample training and evaluation scripts (train.sh, eval.sh).

Maintenance & Community

The project is associated with the authors of the arXiv paper. No specific community channels (Discord, Slack) or active maintenance signals are mentioned in the README.

Licensing & Compatibility

The README does not explicitly state a license. It is presented as an official code release for research purposes. Compatibility for commercial use or closed-source linking is not specified.

Limitations & Caveats

The reported performance metrics may have slight variations (+-0.1 IOU) due to training run variance. The project relies on the nuScenes dataset, and setup requires careful data preparation. The absence of an explicit license may pose a barrier for commercial adoption.