jdc08161063 / 3dlocalmultiviewdesc

Repository for "End-to-End Learning Local Multi-view Descriptors for 3D Point Clouds"

End-to-End Learning Local Multi-view Descriptors for 3D Point Clouds

By Lei Li, Siyu Zhu, Hongbo Fu, Ping Tan, and Chiew-Lan Tai. (CVPR 2020)

In this work, we propose an end-to-end framework to learn local multi-view descriptors for 3D point clouds. To adopt a similar multi-view representation, existing studies use hand-crafted viewpoints for rendering in a preprocessing stage, which is detached from the subsequent descriptor learning stage. In our framework, we integrate the multi-view rendering into neural networks by using a differentiable renderer, which allows the viewpoints to be optimizable parameters for capturing more informative local context of interest points. To obtain discriminative descriptors, we also design a soft-view pooling module to attentively fuse convolutional features across views. Extensive experiments on existing 3D registration benchmarks show that our method outperforms existing local descriptors both quantitatively and qualitatively.

Link

Paper

Citation

@InProceedings{Li_2020_CVPR,
    author = {Li, Lei and Zhu, Siyu and Fu, Hongbo and Tan, Ping and Tai, Chiew-Lan},
    title = {End-to-End Learning Local Multi-view Descriptors for 3D Point Clouds},
    booktitle = {The IEEE Conference on Computer Vision and Pattern Recognition (CVPR)},
    year = {2020}
}

Instructions

Dependencies

• CUDA & CUDNN

• Python >= 3.6

• Install packages by

pip install -r requirements.txt

(Or if you are familiar with Docker, a Dockerfile is provided in folder docker for building a Docker image that includes a complete running environment.)

3DMatch Benchmark

Training

Download the 3DMatch RGB-D data from here.

Go to folder data/preprocess. Use the following scripts for preprocessing and generating training data.

• fuse_fragments_3DMatch.py for generating point cloud fragments from the RGB-D data.
• compute_radius.py for computing point radius. (May skip it to save time & space if to use fixed-radius point rendering)
• compute_overlap.py for finding partially overlapped fragment pairs.
• compute_kpt_pairs.py for selecting point pairs in overlapped regions for batch-hard training.

Go to folder scripts. Fill the paths in configs/ours_3dmatch.yaml and run

python main_mvdesc.py train configs/ours_3dmatch.yaml

A copy of the trained weights is located in scripts/ours_3dmatch.

Evaluation

Download the 3DMatch geometric registration benchmark from here. If you use these data in your work, please consider citing [1].

Go to folder scripts. Fill the paths in configs/ours_3dmatch.yaml and then extract the local multi-view descriptors by running

python main_mvdesc.py test configs/ours_3dmatch.yaml

(Or the extracted descriptors can be directly downloaded from here.)

Compute the recall metric by running evaluation/eval_geomreg_3dmatch.sh.

ETH Benchmark

Download the ETH benchmark from here. If you use these data in your work, please consider citing [2, 3].

Go to folder scripts. Fill the paths in configs/ours_eth.yaml and then extract the local multi-view descriptors by running

python main_mvdesc.py test configs/ours_eth.yaml

(Or the extracted descriptors can be directly downloaded from here.)

Compute the recall metric by running evaluation/eval_geomreg_eth.sh.

References

1. Zeng et al. 3DMatch: Learning Local Geometric Descriptors from RGB-D Reconstructions. CVPR 2017.
2. Pomerleau et al. Challenging data sets for point cloud registration algorithms. IJRR, 2012.
3. Gojcic et al. The Perfect Match: 3D Point Cloud Matching with Smoothed Densities. CVPR 2019.
4. Liu et al. Soft Rasterizer: A differentiable renderer for image-based 3d reasoning. ICCV 2019.