PRIME: A few primitives can boost robustness to common corruptions

Abstract

Despite their impressive performance on image classification tasks, deep networks have a hard time generalizing to many common corruptions of their data. To fix this vulnerability, prior works have mostly focused on increasing the complexity of their training pipelines, combining multiple methods, in the name of diversity. However, in this work, we take a step back and follow a principled approach to achieve robustness to common corruptions. We propose PRIME, a general data augmentation scheme that consists of simple families of max-entropy image transformations. We show that PRIME outperforms the prior art for corruption robustness, while its simplicity and plug-and-play nature enables it to be combined with other methods to further boost their robustness. Furthermore, we analyze PRIME to shed light on the importance of the mixing strategy on synthesizing corrupted images, and to reveal the robustnessaccuracy trade-offs arising in the context of common corruptions. Finally, we show that the computational efficiency of our method allows it to be easily used in both on-line and off-line data augmentation schemes 1 .

Data augmentation with mixtures of max-entropy transformations for filling-level classification

“…We investigated the impact of data augmentation on classifying the filling level of a container, when the containers in the test-time images have different properties than those in the training set. We compared transfer learning -with or without adversarial training [13] -and a principled augmentation approach that explicitly operates on the geometry, color and spatial frequencies of the training images [19] in order to generalize the shape, color, and spectral content of an available, limited training dataset. We showed that the principled augmentation can either replace transfer learning approaches, which are computationally more expensive, or be combined with adversarial transfer learning to improve its performance.…”

“…The performance on the validation set of C-CMM on the "shifted" containers is systematically lower than that on containers that share similarities with those in the training set (overfitting) [13]. To address this limitation, we consider a data augmentation scheme that generates diverse augmentations using a set of primitive max-entropy transformations on the spatial τ , color, γ, and spectral, ω, domain [19]. We expect these transformations to relate to the changes we want to introduce during training: container shape through τ , container color through γ, and illumination and texture through ω.…”

“…A transformed image x T is generated through a convex combination of n basic augmentations (width) consisting of the composition of m of its max-entropy transformations (depth) [19]:…”

“…2). For the smoothness parameter K of these transformations, we keep the original values [19]: K γ =500 for the color domain and K ω =3 for the spectral domain. As for the parameter strength, since very strong changes could remove task-specific information in the images, we only slightly manipulate the color of the pixels and the frequency information of the images, and hence we set σ 2 γ =0.001 and σ 2 ω =0.01, respectively.…”

“…In this work, we address the filling-level classification training problem using data augmentation with a family of maxentropy transformations, along with a mixing strategy [19]. These transformations allow us to generate diverse and targeted augmentations that can be tuned to generate samples tailored to train classifiers that generalize on containers with shapes (Fig.…”