Using Videos to Evaluate Image Model Robustness

By Keren Gu, Brandon Yang, Jiquan Ngiam, Quoc V. Le, Jonathon Shlens. 2019

44 citations recorded by the source corpus.

Human visual systems are robust to a wide range of image transformations that are challenging for artificial networks. We present the first study of image model robustness to the minute transformations found across video frames, which we term "natural robustness". Compared to previous studies on adversarial examples and synthetic distortions, natural robustness captures a more diverse set of common image transformations that occur in the natural environment. Our study across a dozen model architectures shows that more accurate models are more robust to natural transformations, and that robustness to synthetic color distortions is a good proxy for natural robustness. In examining brittleness in videos, we find that majority of the brittleness found in videos lies outside the typical definition of adversarial examples (99.9\%). Finally, we investigate training techniques to reduce brittleness and find that no single technique systematically improves natural robustness across twelve tested architectures.

Subfields: Artificial Intelligence, Biophysics, Computer Vision and Pattern Recognition

How researchers see this paper

The researchers whose trusted networks rank it highest, strongest advocate first. Every researcher gets a different order.

In Artificial Intelligence

In Biophysics

In Computer Vision and Pattern Recognition

Citations in the corpus

Measuring Robustness to Natural Distribution Shifts in Image Classification (2020, methodology): “Appendix L.3 speciﬁcally revisits consistency shifts and explains why, in contrast to previous work [35], we ﬁnd consistency robustness is only weakly correlated with color corruption robustness.”
Do Image Classifiers Generalize Across Time? (2019, methodology): “We discard such frame pairs by employing human expert labelers to evaluate model robustness only on perceptually similar images, unlike prior work [12].”
AugMix: A Simple Data Processing Method to Improve Robustness and Uncertainty (2019, background): “Further, networks trained on translation augmentations remain highly sensitive to images shifted by a single pixel (Gu et al., 2019; Hendrycks & Dietterich, 2019).”
Toward Causal Representation Learning (2021, background): “…and detection methods with respect to simple algorithmically generated interventions, such as spatial shifts, blur, changes in brightness or contrast [107], [170], time consistency [95], [227], control over background and rotation [12], as well as images collected in multiple environments [20].”
Towards Causal Representation Learning (2021, background): “were proposed to specifically test generalization of classification and detection methods with respect to simple algorithmically generated interventions like spatial shifts, blur, changes in brightness or contrast [106, 170], time consistency [94, 227], control over background and rotation [11], as ”
Improving Robustness Without Sacrificing Accuracy with Patch Gaussian Augmentation (2019, background): “Research in neural network robustness has tried to quantify the problem by establishing benchmarks that directly measure it [24, 21] and comparing the performance of humans and neural networks [16].”
Noise as a Resource for Learning in Knowledge Distillation (2020, methodology): “[18] measured model’s robustness to the minute transformations found across video frames which they refer to as natural robustness and found state-of-the-art classiﬁers to be brittle to these transformations.”
Data-centric Artificial Intelligence: A Survey (2023, unknown): “To illustrate, some studies [85, 201] generate various sets of contiguous video frames that appear visually similar to humans but lead to inconsistent predictions due to the small perturbations.”
A systematic framework for natural perturbations from videos (2019, methodology): “To the best of our knowledge, there was also no exhaustive human verification of the adversarial frames in [13], while all perturbed images in our dataset were verified by humans to be similar to the corresponding anchor frame.”
The Evolution of Out-of-Distribution Robustness Throughout Fine-Tuning (2021, background): “However, the vast majority of models show a signiﬁcant drop in performance from in-distribution (ID) data to out-of-distribution (OOD) data [28, 33, 3, 40, 12, 1, 36, 8, 43].”
Evaluating Machine Accuracy on ImageNet (2020, background): “Other forms of natural distribution shift such as robustness to temporal changes could be one candidate (Gu et al., 2019; Shankar et al., 2019).”