Paper: MinkLoc++: Lidar and Monocular Image Fusion for Place Recognition accepted for International Joint Conference on Neural Networks (IJCNN) 2021 ArXiv

Jacek Komorowski, Monika Wysoczańska, Tomasz Trzciński

Warsaw University of Technology

• MinkLoc3D: Point Cloud Based Large-Scale Place Recognition (WACV 2021): MinkLoc3D
• Large-Scale Topological Radar Localization Using Learned Descriptors (ICONIP 2021): RadarLoc
• EgonNN: Egocentric Neural Network for Point Cloud Based 6DoF Relocalization at the City Scale (IEEE Robotics and Automation Letters April 2022): EgoNN
• Improving Point Cloud Based Place Recognition with Ranking-based Loss and Large Batch Training (2022): MinkLoc3Dv2

We present a discriminative multimodal descriptor based on a pair of sensor readings: a point cloud from a LiDAR and an image from an RGB camera. Our descriptor, named MinkLoc++, can be used for place recognition, re-localization and loop closure purposes in robotics or autonomous vehicles applications. We use late fusion approach, where each modality is processed separately and fused in the final part of the processing pipeline. The proposed method achieves state-of-the-art performance on standard place recognition benchmarks. We also identify dominating modality problem when training a multimodal descriptor. The problem manifests itself when the network focuses on a modality with a larger overfit to the training data. This drives the loss down during the training but leads to suboptimal performance on the evaluation set. In this work we describe how to detect and mitigate such risk when using a deep metric learning approach to train a multimodal neural network.

If you find this work useful, please consider citing:

@INPROCEEDINGS{9533373,  
   author={Komorowski, Jacek and Wysoczańska, Monika and Trzcinski, Tomasz},  
   booktitle={2021 International Joint Conference on Neural Networks (IJCNN)},   
   title={MinkLoc++: Lidar and Monocular Image Fusion for Place Recognition},   
   year={2021},  
   doi={10.1109/IJCNN52387.2021.9533373}
}

Code was tested using Python 3.8 with PyTorch 1.9.1 and MinkowskiEngine 0.5.4 on Ubuntu 20.04 with CUDA 10.2.

The following Python packages are required:

• PyTorch (version 1.9.1)
• MinkowskiEngine (version 0.5.4)
• pytorch_metric_learning (version 1.0 or above)
• tensorboard
• colour_demosaicing

Modify the PYTHONPATH environment variable to include absolute path to the project root folder:

export PYTHONPATH=$PYTHONPATH:/home/.../MinkLocMultimodal

MinkLoc++ is a multimodal descriptor based on a pair of inputs:

• a 3D point cloud constructed by aggregating multiple 2D LiDAR scans from Oxford RobotCar dataset,
• a corresponding RGB image from the stereo-center camera.

We use 3D point clouds built by authors of PointNetVLAD: Deep Point Cloud Based Retrieval for Large-Scale Place Recognition paper (link). Each point cloud is built by aggregating 2D LiDAR scans gathered during the 20 meter vehicle traversal. For details see PointNetVLAD paper or their github repository (link). You can download training and evaluation point clouds from here (alternative link).

After downloading the dataset, you need to edit config_baseline_multimodal.txt configuration file (in config folder). Set dataset_folder parameter to point to a root folder of PointNetVLAD dataset with 3D point clouds. image_path parameter must be a folder where downsampled RGB images from Oxford RobotCar dataset will be saved. The folder will be created by generate_rgb_for_lidar.py script.

Run the below code to generate training pickles (with positive and negative point clouds for each anchor point cloud) and evaluation pickles. Training pickle format is optimized and different from the format used in PointNetVLAD code.

cd generating_queries/ 

# Generate training tuples for the Baseline Dataset
python generate_training_tuples_baseline.py --dataset_root <dataset_root_path>

# Generate training tuples for the Refined Dataset
python generate_training_tuples_refine.py --dataset_root <dataset_root_path>

# Generate evaluation tuples
python generate_test_sets.py --dataset_root <dataset_root_path>

<dataset_root_path> is a path to dataset root folder, e.g. /data/pointnetvlad/benchmark_datasets/. Before running the code, ensure you have read/write rights to <dataset_root_path>, as training and evaluation pickles are saved there.

RGB images are taken directly from Oxford RobotCar dataset. First, you need to download stereo camera images from Oxford RobotCar dataset. See dataset website for details (link). After downloading Oxford RobotCar dataset, run generate_rgb_for_lidar.py script. The script finds 20 closest RGB images in RobotCar dataset for each 3D point cloud, downsamples them and saves them in the target directory (image_path parameter in config_baseline_multimodal.txt). During the training an input to the network consists of a 3D point cloud and one RGB image randomly chosen from these 20 corresponding images. During the evaluation, a network input consists of a 3D point cloud and one RGB image with the closest timestamp.

cd scripts/ 

# Generate training tuples for the Baseline Dataset
python generate_rgb_for_lidar.py --config ../config/config_baseline_multimodal.txt --oxford_root <path_to_Oxford_RobotCar_dataset>

Alternatively you can download pre-processed and downsampled RobotCar images from this link.

MinkLoc++ can be used in unimodal scenario (3D point cloud input only) and multimodal scenario (3D point cloud + RGB image input). To train MinkLoc++ network, download and decompress the 3D point cloud dataset and generate training pickles as described above. To train the multimodal model (3D+RGB) download the original Oxford RobotCar dataset and extract RGB images corresponding to 3D point clouds as described above. Edit the configuration files:

• config_baseline_multimodal.txt when training a multimodal (3D+RGB) model
• config_baseline.txt and config_refined.txt when train unimodal (3D only) model

Set dataset_folder parameter to the dataset root folder, where 3D point clouds are located. Set image_path parameter to the path with RGB images corresponding to 3D point clouds, extracted from Oxford RobotCar dataset using generate_rgb_for_lidar.py script (only when training a multimodal model). Modify batch_size_limit parameter depending on the available GPU memory. Default limits requires 11GB of GPU RAM.

To train the multimodal model (3D+RGB), run:

cd training

python train.py --config ../config/config_baseline_multimodal.txt --model_config ../models/minklocmultimodal.txt

To train a unimodal model (3D only) model run:

cd training

# Train unimodal (3D only) model on the Baseline Dataset
python train.py --config ../config/config_baseline.txt --model_config ../models/minkloc3d.txt

# Train unimodal (3D only) model on the Refined Dataset
python train.py --config ../config/config_refined.txt --model_config ../models/minkloc3d.txt

Pretrained models are available in weights directory

• minkloc_multimodal.pth multimodal model (3D+RGB) trained on the Baseline Dataset with corresponding RGB images
• minkloc3d_baseline.pth unimodal model (3D only) trained on the Baseline Dataset
• minkloc3d_refined.pth unimodal model (3D only) trained on the Refined Dataset

To evaluate pretrained models run the following commands:

cd eval

# To evaluate the multimodal model (3D+RGB only) trained on the Baseline Dataset
python evaluate.py --config ../config/config_baseline_multimodal.txt --model_config ../models/minklocmultimodal.txt --weights ../weights/minklocmultimodal_baseline.pth

# To evaluate the unimodal model (3D only) trained on the Baseline Dataset
python evaluate.py --config ../config/config_baseline.txt --model_config ../models/minkloc3d.txt --weights ../weights/minkloc3d_baseline.pth

# To evaluate the unimodal model (3D only) trained on the Refined Dataset
python evaluate.py --config ../config/config_refined.txt --model_config ../models/minkloc3d.txt --weights ../weights/minkloc3d_refined.pth

MinkLoc++ performance (measured by Average Recall@1%) compared to the state of the art:

1. Y. Pan et al., "CORAL: Colored structural representation for bi-modal place recognition", preprint arXiv:2011.10934 (2020)
2. Y. Lu et al., "PIC-Net: Point Cloud and Image Collaboration Network for Large-Scale Place Recognition", preprint arXiv:2008.00658 (2020)
3. M. A. Uy and G. H. Lee, "PointNetVLAD: Deep Point Cloud Based Retrieval for Large-Scale Place Recognition", 2018 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR)
4. W. Zhang and C. Xiao, "PCAN: 3D Attention Map Learning Using Contextual Information for Point Cloud Based Retrieval", 2019 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR)
5. Q. Sun et al., "DAGC: Employing Dual Attention and Graph Convolution for Point Cloud based Place Recognition", Proceedings of the 2020 International Conference on Multimedia Retrieval
6. Z. Liu et al., "LPD-Net: 3D Point Cloud Learning for Large-Scale Place Recognition and Environment Analysis", 2019 IEEE/CVF International Conference on Computer Vision (ICCV)
7. L. Hui et al., "Efficient 3D Point Cloud Feature Learning for Large-Scale Place Recognition" preprint arXiv:2101.02374 (2021)
8. Y. Xia et al., "SOE-Net: A Self-Attention and Orientation Encoding Network for Point Cloud based Place Recognition", 2021 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR)
9. J. Komorowski, "MinkLoc3D: Point Cloud Based Large-Scale Place Recognition", Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision (WACV), (2021)
10. Z. Zhou et al., "NDT-Transformer: Large-scale 3D Point Cloud Localisation Using the Normal Distribution Transform Representation", 2021 IEEE International Conference on Robotics and Automation (ICRA)

• J. Komorowski, M. Wysoczanska, T. Trzcinski, "MinkLoc++: Lidar and Monocular Image Fusion for Place Recognition", accepted for International Joint Conference on Neural Networks (IJCNN), (2021)

Our code is released under the MIT License (see LICENSE file for details).