Projecting Your View Attentively: Monocular Road Scene Layout Estimation via Cross-view Transformation

Weixiang Yang, Qi Li, Wenxi Liu, Yuanlong Yu, Yuexin Ma, Shengfeng He, Jia Pan

Paper

Accepted to CVPR 2021

Abstract

HD map reconstruction is crucial for autonomous driving. LiDAR-based methods are limited due to the deployed expensive sensors and time-consuming computation. Camera-based methods usually need to separately perform road segmentation and view transformation, which often causes distortion and the absence of content. To push the limits of the technology, we present a novel framework that enables reconstructing a local map formed by road layout and vehicle occupancy in the bird's-eye view given a front-view monocular image only. In particular, we propose a cross-view transformation module, which takes the constraint of cycle consistency between views into account and makes full use of their correlation to strengthen the view transformation and scene understanding. Considering the relationship between vehicles and roads, we also design a context-aware discriminator to further refine the results. Experiments on public benchmarks show that our method achieves the state-of-the-art performance in the tasks of road layout estimation and vehicle occupancy estimation. Especially for the latter task, our model outperforms all competitors by a large margin. Furthermore, our model runs at 35 FPS on a single GPU, which is efficient and applicable for real-time panorama HD map reconstruction.

Contributions

We propose a novel framework that reconstructs a local map formed by top-view road scene layout and vehicle occupancy using a single monocular front-view image only. In particular, we propose a cross-view transformation module which leverages the cycle consistency between views and their correlation to strengthen the view transformation.
We also propose a context-aware discriminator that considers the spatial relationship between vehicles and roads in the task of estimating vehicle occupancies.
On public benchmarks, it is demonstrated that our model achieves the state-of-the-art performance for the tasks of road layout and vehicle occupancy estimation.

Approach overview

Repository Structure

cross-view/
├── crossView            # Contains scripts for dataloaders and network/model architecture
└── datasets             # Contains datasets
    ├── argoverse        # argoverse dataset
    ├── kitti            # kitti dataset 
├── log                  # Contains a log of network/model
├── losses               # Contains scripts for loss of network/model
├── models               # Contains the saved model of the network/model
├── output               # Contains output of network/model
└── splits
    ├── 3Dobject         # Training and testing splits for KITTI 3DObject Detection dataset 
    ├── argo             # Training and testing splits for Argoverse Tracking v1.0 dataset
    ├── odometry         # Training and testing splits for KITTI Odometry dataset
    └── raw              # Training and testing splits for KITTI RAW dataset(based on Schulter et. al.)

Installation

We recommend setting up a Python 3.7 and Pytorch 1.0 Virtual Environment and installing all the dependencies listed in the requirements file.

git clone https://github.com/JonDoe-297/cross-view.git

cd cross-view
pip install -r requirements.txt

Datasets

In the paper, we've presented results for KITTI 3D Object, KITTI Odometry, KITTI RAW, and Argoverse 3D Tracking v1.0 datasets. For comparison with Schulter et. al., We've used the same training and test splits sequences from the KITTI RAW dataset. For more details about the training/testing splits one can look at the splits directory. And you can download Ground-truth from Monolayout.

# Download KITTI RAW
./data/download_datasets.sh raw

# Download KITTI 3D Object
./data/download_datasets.sh object

# Download KITTI Odometry
./data/download_datasets.sh odometry

# Download Argoverse Tracking v1.0
./data/download_datasets.sh argoverse

The above scripts will download, unzip and store the respective datasets in the datasets directory.

datasets/
└── argoverse                          # argoverse dataset
    └── argoverse-tracking
        └── train1
            └── 1d676737-4110-3f7e-bec0-0c90f74c248f
                ├── car_bev_gt         # Vehicle GT
                ├── road_gt            # Road GT
                ├── stereo_front_left  # RGB image
└── kitti                              # kitti dataset 
    └── object                         # kitti 3D Object dataset 
        └── training
            ├── image_2                # RGB image
            ├── vehicle_256            # Vehicle GT
    ├── odometry                       # kitti odometry dataset 
        └── 00
            ├── image_2                # RGB image
            ├── road_dense128  # Road GT
    ├── raw                            # kitti raw dataset 
        └── 2011_09_26
            └── 2011_09_26_drive_0001_sync
                ├── image_2            # RGB image
                ├── road_dense128      # Road GT

Training

Prepare the corresponding dataset
Run training

# Corss view Road (KITTI Odometry)
python3 train.py --type static --split odometry --data_path ./datasets/odometry/ --model_name <Model Name with specifications>

# Corss view Vehicle (KITTI 3D Object)
python3 train.py --type dynamic --split 3Dobject --data_path ./datasets/kitti/object/training --model_name <Model Name with specifications>

# Corss view Road (KITTI RAW)
python3 train.py --type static --split raw --data_path ./datasets/kitti/raw/  --model_name <Model Name with specifications>

# Corss view Vehicle (Argoverse Tracking v1.0)
python3 train.py --type dynamic --split argo --data_path ./datasets/argoverse/ --model_name <Model Name with specifications>

# Corss view Road (Argoverse Tracking v1.0)
python3 train.py --type static --split argo --data_path ./datasets/argoverse/ --model_name <Model Name with specifications>

The training model are in "models" (default: ./models)

Testing

Download pre-trained models
Run testing

python3 test.py --type <static/dynamic> --model_path <path to the model directory> --image_path <path to the image directory>  --out_dir <path to the output directory>

The results are in "output" (default: ./output)

Evaluation

Prepare the corresponding dataset
Download pre-trained models
Run evaluation

# Evaluate on KITTI Odometry 
python3 eval.py --type static --split odometry --model_path <path to the model directory> --data_path ./datasets/odometry --height 512 --width 512 --occ_map_size 128

# Evaluate on KITTI 3D Object
python3 eval.py --type dynamic --split 3Dobject --model_path <path to the model directory> --data_path ./datasets/kitti/object/training

# Evaluate on KITTI RAW
python3 eval.py --type static --split raw --model_path <path to the model directory> --data_path ./datasets/kitti/raw/

# Evaluate on Argoverse Tracking v1.0 (Road)
python3 eval.py --type static --split argo --model_path <path to the model directory> --data_path ./datasets/kitti/argoverse/

# Evaluate on Argoverse Tracking v1.0 (Vehicle)
python3 eval.py --type dynamic --split argo --model_path <path to the model directory> --data_path ./datasets/kitti/argoverse

The results are in "output" (default: ./output)

Pretrained Models

The following table provides links to the pre-trained models for each dataset mentioned in our paper. The table also shows the corresponding evaluation results for these models.

Dataset	Segmentation Objects	mIOU(%)	mAP(%)	Pretrained Model
KITTI 3D Object	Vehicle	38.85	51.04	link
KITTI Odometry	Road	77.47	86.39	link
KITTI Raw	Road	68.26	79.65	link
Argoverse Tracking	Vehicle	47.87	62.69	link
Argoverse Tracking	Road	76.56	87.30	link

Results

Contact

If you meet any problems, please describe them in issues or contact:

Weixiang Yang: [email protected]

License

This project is released under the MIT License (refer to the LICENSE file for details).This project partially depends on the sources of Monolayout

[CVPR'21] Projecting Your View Attentively: Monocular Road Scene Layout Estimation via Cross-view Transformation

Related tags

Overview

Projecting Your View Attentively: Monocular Road Scene Layout Estimation via Cross-view Transformation

Weixiang Yang, Qi Li, Wenxi Liu, Yuanlong Yu, Yuexin Ma, Shengfeng He, Jia Pan

Paper

Accepted to CVPR 2021

Abstract

Contributions

Approach overview

Repository Structure

Installation

Datasets

Training

Testing

Evaluation

Pretrained Models

Results

Contact

License

Owner

SimulLR - PyTorch Implementation of SimulLR

Official code for "On the Frequency Bias of Generative Models", NeurIPS 2021

Pytorch Implementation of Various Point Transformers

Codebase for INVASE: Instance-wise Variable Selection - 2019 ICLR

Detecting Potentially Harmful and Protective Suicide-related Content on Twitter

Official implementation of our neural-network-based fast diffuse room impulse response generator (FAST-RIR)

Auto-Encoding Score Distribution Regression for Action Quality Assessment

Contrastive unpaired image-to-image translation, faster and lighter training than cyclegan (ECCV 2020, in PyTorch)

A PyTorch implementation for our paper "Dual Contrastive Learning: Text Classification via Label-Aware Data Augmentation".

meProp: Sparsified Back Propagation for Accelerated Deep Learning

[CVPR'22] COAP: Learning Compositional Occupancy of People

This is just a funny project that we want to see AutoEncoder (AE) can actually work to enhance the features we want

Head2Toe: Utilizing Intermediate Representations for Better OOD Generalization

Harmonic Memory Networks for Graph Completion

Temporally Efficient Vision Transformer for Video Instance Segmentation, CVPR 2022, Oral

Implementation of Monocular Direct Sparse Localization in a Prior 3D Surfel Map (DSL)

Pytorch implementation for "Distribution-Balanced Loss for Multi-Label Classification in Long-Tailed Datasets" (ECCV 2020 Spotlight)

Practical tutorials and labs for TensorFlow used by Nvidia, FFN, CNN, RNN, Kaggle, AE

Experimental solutions to selected exercises from the book [Advances in Financial Machine Learning by Marcos Lopez De Prado]

DeepDiffusion: Unsupervised Learning of Retrieval-adapted Representations via Diffusion-based Ranking on Latent Feature Manifold