Neural Networks have been shown to be sensitive to common perturbations such as blur, Gaussian noise, rotations, etc. They are also vulnerable to some artificial malicious corruptions called adversarial examples. The adversarial examples study has recently become very popular and it sometimes even reduces the term "adversarial robustness" to the term "robustness". Yet, we do not know to what extent the adversarial robustness is related to the global robustness. Similarly, we do not know if a robustness to various common perturbations such as translations or contrast losses for instance, could help with adversarial corruptions. We intend to study the links between the robustnesses of neural networks to both perturbations. With our experiments, we provide one of the first benchmark designed to estimate the robustness of neural networks to common perturbations. We show that increasing the robustness to carefully selected common perturbations, can make neural networks more robust to unseen common perturbations. We also prove that adversarial robustness and robustness to common perturbations are independent. Our results make us believe that neural network robustness should be addressed in a broader sense.
As difficult vision-based tasks like object detection and monocular depth estimation are making their way in real-time applications and as more light weighted solutions for autonomous vehicles navigation systems are emerging, obstacle detection and collision prediction are two very challenging tasks for small embedded devices like drones. We propose a novel light weighted and time-efficient vision-based solution to predict Time-to-Collision from a monocular video camera embedded in a smartglasses device as a module of a navigation system for visually impaired pedestrians. It consists of two modules: a static data extractor made of a convolutional neural network to predict the obstacle position and distance and a dynamic data extractor that stacks the obstacle data from multiple frames and predicts the Time-to-Collision with a simple fully connected neural network. This paper focuses on the Time-to-Collision network’s ability to adapt to new sceneries with different types of obstacles with supervised learning.
The main goal of this paper is to study Image Aesthetic Assessment (IAA) indicating images as high or low aesthetic. The main contributions concern three points. Firstly, following the idea that photos in different categories (human, flower, animal, landscape, …) are taken with different photographic rules, image aesthetic should be evaluated in a different way for each image category. Large field images and close-up images are two typical categories of images with opposite photographic rules so we want to investigate the intuition that prior Large field/Close-up Image Classification (LCIC) might improve the performance of IAA. Secondly, when a viewer looks at a photo, some regions receive more attention than other regions. Those regions are defined as Regions Of Interest (ROI) and it might be worthy to identify those regions before IAA. The question “Is it worthy to extract some ROIs before IAA?” is considered by studying Region Of Interest Extraction (ROIE) before investigating IAA based on each feature set (global image features, ROI features and background features). Based on the answers, a new IAA model is proposed. The last point is about a comparison between the efficiency of handcrafted and learned features for the purpose of IAA.
11 January 2013 in Multimedia Tools and Applications
The main purpose of this paper is the study of the correlations between Image Aesthetic (IA) and Image Naturalness (IN) and the analysis of the influence of IA and IN on Image Quality (IQ) in different contexts. The first contribution is a study about the potential relationships between IA and IN. For that study, two sub-questions are considered. The first one is to validate the idea that IA and IN are not correlated to each other. The second one is about the influence of IA and IN features on Image Naturalness Assessment (INA) and Image Aesthetic Assessment (IAA), respectively. Secondly, it is obvious that IQ is related to IA and IN, but the exact influence of IA and IN on IQ has not been evaluated. Besides that, the context impact on those influences has not been clarified, so the second contribution is to investigate the influence of IA and IN on IQ in different contexts. The results obtained from rigorous experiments prove that although there are moderate and weak correlations between IA and IN, they are still two different components of IQ. It also appears that viewers’ IQ perception is affected by some contextual factors, and the influence of IA and IN on IQ depends on the considered context.