News

• October 14, 2022: We've released code & models for map-view segmentation

• June 9, 2022: We've released the tech report for Geometry-guided Kernel Transformer (GKT). This work is still in progress and code/models are coming sonn. Please stay tuned! ☕️

Introduction

We present a novel and efficient 2D-to-BEV transformation, Geometry-guided Kernel Transformer (GKT).

• GKT leverages geometric priors to guide the transformers to focus on discriminative regions for generating BEV representation with surrouding-view image features.
• GKT is based on kernel-wise attention and much efficient, especially with LUT indexing.
• GKT is robust to the deviation of cameras, making the 2D-to-BEV transformation more stable and reliable.

Getting Started

git clone https://github.com/hustvl/GKT.git

Map-view nuScenes Segmentation

Models

Note: FPS are measured on one 2080 Ti GPU.

Usage

For map-view nuScenes segmentation, we mainly build the GKT based on the awesome CrossViewTransformer.

# map-view segmentation
cd segmentation

Prerequisites

# install dependencies
pip install -r reuqirements.txt
pip install -e .

Preparing the Dataset

• Dataset setup
• Label generation

Training / Testing / Benchmarking

• Pretrained model

Download the pretrained model efficientnet-b4-6ed6700e.pth

mkdir pretrained_models
cd pretrained_models
# place the pretrained model here

• Training

python scripts/train.py +experiment=gkt_nuscenes_vehicle_kernel_7x1.yaml  data.dataset_dir=<path/to/nuScenes> data.labels_dir=<path/to/labels>

• Testing

Using the absolute path of the checkpoint is better.

python scripts/eval.py +experiment=gkt_nuscenes_vehicle_kernel_7x1.yaml data.dataset_dir=<path/to/nuScenes> data.labels_dir=<path/to/labels> experiment.ckptt <path/to/checkpoint>

• Evalutating Speed

python scripts/speed.py +experiment=gkt_nuscenes_vehicle_kernel_7x1.yaml data.dataset_dir=<path/to/nuScenes> data.labels_dir=<path/to/labels>

3D Object Detection

coming soon.

Acknowledgements

We sincerely appreciate the awesome repos cross_view_transformers and fiery!

License

GKT is released under the MIT Licence.

Citation

If you find GKT is useful in your research or applications, please consider giving us a star 🌟 and citing it by the following BibTeX entry.

@article{GeokernelTransformer,
  title={Efficient and Robust 2D-to-BEV Representation Learning via Geometry-guided Kernel Transformer},
  author={Chen, Shaoyu and Cheng, Tianheng and Wang, Xinggang and Meng, Wenming and Zhang, Qian and Liu, Wenyu},
  journal={arXiv preprint arXiv:2206.04584},
  year={2022}
}