Given a point set \(\{\mathbf{p}_i \}_{i=1}^n\) with \(\mathbf{p}_i \in \mathbb{R}^3\) represented by a matrix \(\mathbf{x}_0 \in \mathbb{R}^{n \times 3}\). To maintain the generality of our approach, we assume that the input \(\mathbf{x}_0\) is passed to pre-trained network \(f\) for feature extraction, and the features are passed to a network \(g\) for final label prediction \(f(\mathbf{x}_0)\) throughout the inference process. Our goal is to achieve performance improvement in the downstream task via test-time augmentation, where the final prediction can be defined as: \[g(\phi(f(\mathbf{x}_0), f(\mathbf{x}_1), f(\mathbf{x}_2), ...))\] where \(\phi\) is an aggregation function to combine multiple features resulting from the original point set \(\mathbf{x}_0\) and the augmented point sets \(\mathbf{x}_1\), \(\mathbf{x}_2\), etc. Note that the network \(f\) and \(g\) are pre-trained and left untouched in test-time augmentation; only the input is augmented. Traditionally, a simple method for test-time augmentation is jittering, which adds Gaussian noise to perturb the point cloud \(\mathbf{x}_0\) to generate an augmented point cloud \(\mathbf{x}_k\): \[\mathbf{x}_k = \mathbf{x}_0 + \lambda \mathbf{z}_k\] where \(\mathbf{z}_k \sim \mathcal{N}(0, I)\) is a random noise vector from a normal distribution, and \(\lambda\) is a scalar value to control the noise magnitude. This simple augmentation has been widely adopted since the seminal PointNet. An issue of such augmentation is that it does not consider the underlying surface or point distribution because the noise \(\mathbf{z}\) is independent of \(\mathbf{x}_0\), resulting in marginal performance improvement in many cases. In this work, we viewpoint set \(\mathbf{x}_0\) as a noisy estimate of a latent surface representation \(\mathcal{S}\), and therefore, we define point cloud augmentation as the process of sampling additional point clouds \(\mathbf{x}_k\) that explain the same surface. We propose to sample augmented point clouds \(\mathbf{x}_k\) (\(k \geq 1\)) in two ways: surface sampling and point cloud up-sampling. The sampled point clouds can then be leveraged for downstream tasks such as classification and segmentation.