MapPrior

ICCV 2023

MapPrior: Bird’s-Eye View Map Layout Estimation with Generative Models
Xiyue Zhu, Vlas Zyrianov, Zhijian Liu, Shenlong Wang

tl;dr MapPrior is a novel BEV perception framework that combines a traditional discriminative BEV perception model with a learned generative model for semantic map layouts. It delivers predictions with better accuracy, realism and uncertainty awareness than its baselines.

Paper | BibTeX | Project Page

Overview of pretrained models

The following table provides an overview of our models.

Model	Modality	mIoU	MMD
MapPrior_L	L	51.1	35.8
MapPrior_C	C	56.7	28.4
MapPrior_CL	C+L	63.1	28.0
MapPrior_L_one_step	L	50.0	50.2
MapPrior_C_one_step	C	56.7	28.7
MapPrior_CL_one_step	C+L	63.0	28.1

Requirements

The code is based on VQGAN and BEVFusion and requires the following libraries:

Python >= 3.8, <3.9
OpenMPI = 4.0.4 and mpi4py = 3.0.3 (Needed for torchpack)
Pillow = 8.4.0 (see here)
PyTorch >= 1.9, <= 1.10.2
torchpack
mmcv==1.4.0
mmcv-full==1.4.0
mmdet==2.20.0
nuscenes-dev-kit
numba==0.48.0
numpy==1.23
tqdm
albumentations==0.4.3
opencv-python==4.1.2.30
pudb==2019.2
imageio==2.9.0
imageio-ffmpeg==0.4.2
pytorch-lightning==1.0.8
omegaconf==2.0.0
test-tube>=0.7.5
streamlit>=0.73.1
einops==0.3.0
more-itertools>=8.0.0
transformers

After installing these dependencies, run the following commands to install MapPrior, a modified version of BEVFusion (under bev_lib folder), and to download pretrained MapPrior and BEVFusion models:

pip install -v -e .
cd bev_lib 
pip install -v -e .
cd ..
./scripts/download_pretrained.sh
./scripts/download_pretrained_bevfusion.sh

Data Preparation

nuScenes

Download nuscenes to MapPrior/data.

Next, prepare the dataset by running (see this link for more details):

python scripts/create_data.py nuscenes --root-path ./data/nuscenes --out-dir ./data/nuscenes --extra-tag nuscenes

After data preparation, you will have the following directory structure:

MapPrior
├── data
│   ├── nuscenes
│   │   ├── maps
│   │   ├── samples
│   │   ├── sweeps
│   │   ├── v1.0-test
│   |   ├── v1.0-trainval
│   │   ├── nuscenes_database
│   │   ├── nuscenes_infos_train.pkl
│   │   ├── nuscenes_infos_val.pkl
│   │   ├── nuscenes_infos_test.pkl
│   │   ├── nuscenes_dbinfos_train.pkl

Saved maps [Optional]

We save nuscenes maps instead of rendering it on the fly using nuscenes data loader to accelerate the training. Please download the saved maps for both training set and validation set and put it under MapPrior/data. This is needed only when training a VQGAN autoencoder from scratch. Alternatively, we provide a script to download all the checkpoint files using wget.

./download_saved_map.sh

Running pretrained models

Please download pretrained models using this link . Alternatively, we provide a script to download all the checkpoint files using wget.

./scripts/download_pretrained.sh

Then use this command to run inference and save the results to a folder

python ./scripts/transformer.py --base [config_path] --resume [checkpoint_path] --save [folder_to_save_results] -t True --gpus 0,

For example, to save the results for MapPrior-L under the folder MapPrior_L_results using the downloaded checkpoint:

python ./scripts/transformer.py --base configs/cond_feature_L.yaml --resume pretrained_models/MapPrior_L --save MapPrior_L_results -t True --gpus 0,

To calculate IOU and MMD, run

python ./scripts/test_metrics.py -s [folder_to_save_results]

Here [folder_to_save_results] is the folder where you saved your results in the last command, e.g. "MapPrior_L_results".

Training models

To train the first stage model, an autoencoder built on the ground truth map, use this command. (Use --gpus flags when using multi-gpu. For example, use "--gpus 0,1,2,3" when using 4 gpus)

python ./scripts/train_autoencoder.py --base configs/custom_vqgan.yaml -t True --gpus 0,

The model file and the training logs will be saevd in MapPrior/logs/custom_vqgan. To use the model in the second stage, move the logs to pretrained models or modify the path in the second stage config file.

mv logs/custom_vqgan pretrained_models/autoencoder

To train the second stage models from scrath, use this command:

python ./scripts/transformer.py --base [config_files] -t True --gpus 0,

For example:

python ./scripts/transformer.py --base configs/cond_feature_L.yaml -t True --gpus 0,

Again, the model checkpoint is saved under logs, to save results using the second stage model, run the same command from Running pretrained models but change the checkpoint_path to the saved checkpoint.

For example,

python ./scripts/transformer.py --base configs/cond_feature_L.yaml --resume pretrained_models/cond_feature_L --save MapPrior_L_results -t True --gpus 0,

Acknowledgements

The architecture of our Model is inspired by VQGAN and BEVFusion

BibTeX

@misc{zhu2023mapprior,
      title={MapPrior: Bird’s-Eye View Perception with Generative Models}, 
      author={Xiyue Zhu and Vlas Zyrianov and Zhijian Liu and Shenlong Wang},
      year={2023},
      booktitle={ICCV},
}

xiyuez2/MapPrior