EdgeGaussians -- 3D Edge Mapping via Gaussian Splatting

Kunal Chelani,Assia Benbihi,Torsten Sattler,Fredrik Kahl

2024-09-20

Abstract:With their meaningful geometry and their omnipresence in the 3D world, edges are extremely useful primitives in computer vision. 3D edges comprise of lines and curves, and methods to reconstruct them use either multi-view images or point clouds as input. State-of-the-art image-based methods first learn a 3D edge point cloud then fit 3D edges to it. The edge point cloud is obtained by learning a 3D neural implicit edge field from which the 3D edge points are sampled on a specific level set (0 or 1). However, such methods present two important drawbacks: i) it is not realistic to sample points on exact level sets due to float imprecision and training inaccuracies. Instead, they are sampled within a range of levels so the points do not lie accurately on the 3D edges and require further processing. ii) Such implicit representations are computationally expensive and require long training times. In this paper, we address these two limitations and propose a 3D edge mapping that is simpler, more efficient, and preserves accuracy. Our method learns explicitly the 3D edge points and their edge direction hence bypassing the need for point sampling. It casts a 3D edge point as the center of a 3D Gaussian and the edge direction as the principal axis of the Gaussian. Such a representation has the advantage of being not only geometrically meaningful but also compatible with the efficient training optimization defined in Gaussian Splatting. Results show that the proposed method produces edges as accurate and complete as the state-of-the-art while being an order of magnitude faster. Code is released at <a class="link-external link-https" href="https://github.com/kunalchelani/EdgeGaussians" rel="external noopener nofollow">this https URL</a>.

Computer Vision and Pattern Recognition

What problem does this paper attempt to address?

### What problems does this paper attempt to solve?

The paper titled "EdgeGaussians - 3D Edge Mapping via Gaussian Point Drawing" aims to solve two main problems in existing 3D edge reconstruction methods:

1. **Sampling accuracy problem**:
- Existing methods based on implicit representations (such as neural implicit fields) need to sample 3D edge points on specific level sets (usually 0 or 1). However, due to floating - point precision limitations and training inaccuracies, it is unrealistic to sample these points precisely. Therefore, in practice, these points are usually sampled within a range of level sets, resulting in these points not accurately falling on the 3D edges, and thus requiring additional post - processing steps to correct these errors.

2. **Computational efficiency problem**:
- These implicit representation methods are computationally expensive during the training process and require a long training time. For example, some of the latest methods may require between 1.5 hours and 14 hours to train on a simple CAD model dataset.

To solve these problems, this paper proposes a new 3D edge mapping method called **EdgeGaussians**. This method simplifies and improves the efficiency of 3D edge reconstruction in the following ways:

- **Explicitly learn 3D edge points and their directions**: Directly learn 3D edge points and their directions without the need for complex level - set sampling and subsequent processing through implicit representations.
- **Represent edge points using 3D Gaussian distributions**: Represent each 3D edge point as the center of a 3D Gaussian distribution, and define the edge direction as the main axis of the Gaussian distribution. This representation is not only geometrically meaningful but also can be quickly trained using efficient training optimization methods (such as the optimization methods in Gaussian point drawing).
- **Significantly improve speed**: Experimental results show that the performance of this method is comparable to existing learning - based methods, but the running speed is an order of magnitude faster.

Through these improvements, the EdgeGaussians method can significantly improve computational efficiency while maintaining reconstruction accuracy.