Article directory
1. Four questions
- What problem to solve
Introduce a data augmentation approach to the point cloud - How to solve
mixup combined point cloud - How does the effect
increase by 1-2 points, which is obvious when the amount of data is small - What's the problem
?
2. The content of the paper
Abstract
This paper presents a method for data augmentation of point clouds using inter-instance interpolation. Interpolation data augmentation is a simple and effective method in the image domain. However, such blending cannot be directly transferred to point clouds, since we do not have a one-to-one correspondence between points of two different objects. This paper defines data augmentation between point clouds as shortest path linear interpolation. To this end, we introduce PointMixup. This interpolation method generates new examples by optimizing the assignment of the path function between two point clouds. We demonstrate that our PointMixup is able to find the shortest path between two point clouds and that the interpolation is assignment-invariant and linear. PointMixup allows the introduction of strong interpolation-based regularizers in the point cloud domain. In experiments, we demonstrate the potential of PointMixup in point cloud classification, especially in the absence of instances, and we simultaneously demonstrate the robustness of PointMixup in enhancing the response to noise and geometric transformations of points.
1 Introduction
mixup can improve the regularization and generalization ability of the model, such as the following work
Several works have shown that generating new training examples, by interpolating images and their corresponding labels, leads to improved network regularization and generalization, e.g., [8, 24, 34, 26]
It is easy to apply mixup in the image field, but it is difficult to apply mixup to point clouds: there is no one-to-one correspondence of coordinates (there are on 2D images, for example, each pixel corresponds to a label)
Such a mixup is feasible in the image domain, due to the regular structure of images and one-to-one correspondences between pixels. However, this setup does not generalize to the point cloud domain, since there is no one-to-one correspondence and ordering between points.
2 Related Work
Deep learning for point clouds.
PointNet++ multi-level is the key (trick)
PointNet++ [16] extends this idea further with hierarchical structure by relying on a heuristic method of farthest point sampling and grouping to build the hierarchy.
Mixup is a model-independent method, which is a data augmentation, simple and effective, and can reduce the generalization error.
To that end, we propose a new model-agnostic data augmentation. We propose a Mixup regularization for point clouds and show that it can build on various architectures to obtain better classification results by reducing the generalization error in classification tasks.
Most relevant work: Li et al. [11]
A very recent work by Li et al. [11] also considers improving point cloud classification by augmentation. They rely on auto-augmentation and a complicated adversarial training procedure, whereas in this work we propose to augment point clouds by interpolation.
Interpolation-based regularization.
Using regularization methods is beneficial to improve generalization ability, which has become standard practice in deep learning
Employing regularization approaches for training deep neural networks to improve their generalization performances have become standard practice in deep learning.
Mixup 和 Manifold Mixup
Recent works consider a regularization by interpolating the example and label pairs, commonly known as Mixup [24, 8, 34]. Manifold Mixup [26] extends Mixup by interpolating the hidden representations at multiple layers.
3 Point cloud augmentation by interpolation
3.1 Problem setting
A simple introduction to mixup, it is easy to use on images because the data is already aligned, but the point cloud is unordered and permutation invariant, and an interpolation method must be found on the unordered set.
3.3 PointMixup: Optimal assignment interpolation for point clouds
? I don't understand, this section...
3.5 Manifold PointMixup: Interpolate between latent point features
Manifold mixup: mixup features
some experimental results
The effect of small amount of data is obvious
3. References
Paper citation : https://paperswithcode.com/paper/pointmixup-augmentation-for-point-clouds
Tips for CNN image classification (1): mixup data enhancement
4. Harvest
As a data enhancement method, mixup can improve 1-2 points. It is more obvious for the small amount of data and can be used as a trick. It can enhance generalization and robustness (?),
but I still don't understand how to apply it to the point cloud..., this part of the pit may depend on the code combination, but I don't understand the formula.
