OOD-CV: A Benchmark for Robustness to Out-of-Distribution Shifts of Individual Nuisances in Natural Images

Abstract

Enhancing the robustness of vision algorithms in real-world scenarios is challenging. One reason is that existing robustness benchmarks are limited, as they either rely on synthetic data or ignore the effects of individual nuisance factors. We introduce OOD-CV-v2 , a benchmark dataset that includes out-of-distribution examples of 10 object categories in terms of pose, shape, texture, context and the weather conditions, and enables benchmarking of models for image classification, object detection, and 3D pose estimation. In addition to this novel dataset, we contribute extensive experiments using popular baseline methods, which reveal that: 1) Some nuisance factors have a much stronger negative effect on the performance compared to others, also depending on the vision task. 2) Current approaches to enhance robustness have only marginal effects, and can even reduce robustness. 3) We do not observe significant differences between convolutional and transformer architectures. We believe our dataset provides a rich test bed to study robustness and will help push forward research in this area. Our dataset can be accessed from http://www.ood-cv.org/challenge.html.

Red Teaming for Multimodal Large Language Models: A Survey

“…They show feasibility of this attack on MiniGPT-4 [17], Fuyu [40], LLaVA [16]. This attack is evaluated on Safety Evaluation Benchmark, focusing on two evaluation scenarios: (a) OODCV-VQA and Counterfactual Variant: A novel VQA dataset is proposed grounded on images from OODCV [41]. The dataset includes questions with pre-defined templates for yes/no or digit responses.…”

Robustness Assessment of AI-Based 2D Object Detection Systems: A Method and Lessons Learned from Two Industrial Cases

“…Regarding natural inputs, several methods have been proposed for defining perturbations and generating perturbed input images, guided by knowledge of the application domain [58], coverage metrics [59,60], or properties that the system must satisfy, such as metamorphic relations [22,61]. In addition, several perturbation benchmarks have been proposed in the literature to assess robustness [7,[62][63][64]. However, few of these methods have been developed and evaluated in the context of 2D object detection systems.…”

“…It aims to verify that inserting objects into the background of an image does not change the results for the other objects. For their part, Zhao et al [63] proposed a natural perturbation benchmark for testing models in the field of computer vision, including object detection models.…”