Convolutional Neural Networks (CNNs) are one of the most commonly used architectures for image-related deep learning studies. Despite its popularity, CNNs have some intrinsic limitations such as losing some of the spatial information and not being robust to affine transformations due to pooling operations. On the other hand, Capsule Networks are composed of groups of neurons, and with the help of its novel routing algorithms, they have the capability for learning high dimensional pose configuration of the objects as well. In this study, we investigate the performance of brand-new Capsule Networks using dynamic routing algorithm on the clothing classification task. To achieve this, we propose 4-layer stacked-convolutional Capsule Network architecture (FashionCapsNet), and train this model on DeepFashion dataset that contains 290k clothing images over 46 different categories. Thereafter, we compare the category classification results of our proposed design and the other state-of-the-art CNN-based methods trained on DeepFashion dataset. As a result of the experimental study, FashionCapsNet achieves 83.81% top-3 accuracy, and 89.83% top-5 accuracy on the clothing classification. Based upon these figures, FashionCapsNet clearly outperforms the earlier methods that neglect pose configuration, and has comparable performance to the baseline study that utilizes an additional landmark information to recover pose configuration. Finally, in the future, proposed FashionCapsNet may inherit extra performance boost on the clothing classification due to advances in the relatively new Capsule Network research.

Test oracles differentiate between the correct and incorrect system behavior. Automation of test oracles for visual output systems mainly involves image comparison, where a snapshot of the output is compared with respect to a reference image. Hereby, the captured snapshot can be subject to variations such as scaling and shifting. These variations lead to incorrect evaluations. Existing approaches employ computer vision techniques to address a specific set of variations. In this article, we introduce ADVISOR, an adjustable framework for test oracle automation of visual output systems. It allows the use of a flexible combination and configuration of computer vision techniques. We evaluated a set of valid configurations with a benchmark dataset collected during the tests of commercial digital TV systems. Some of these configurations achieved up to 3% better overall accuracy with respect to state-of-the-art tools. Further, we observed that there is no configuration that reaches the best accuracy for all types of image variations. We also empirically investigated the impact of significant parameters. One of them is a threshold regarding image matching score that determines the final verdict. This parameter is automatically tuned by offline training. We evaluated runtime performance as well. Results showed that differences among the ADVISOR configurations and state-of-the-art tools are in the order of seconds per image comparison.

White balance (WB) correction is critical for accurate color reproduction in digital images, especially under complex, multi-illuminant lighting conditions. Traditional methods, such as the Gray-World assumption, rely on global statistics and struggle in real-world, non-uniform lighting scenarios. Modern deep learning approaches, including convolutional and attention-based architectures, have significantly advanced WB correction but often fail to explicitly account for higher-order feature distribution statistics, which may limit their robustness in challenging environments. This study introduces a novel framework that leverages Exact Feature Distribution Matching (EFDM) as a loss objective to align feature distributions across multiple moments, including mean, variance, skewness, and kurtosis. By modeling lighting as a style factor, the method explicitly addresses distributional shifts caused by complex illumination, offering a robust solution for WB correction. The framework integrates EFDM with a Vision Transformer architecture, enabling precise handling of global and local lighting variations. Extensive experiments on the large-scale multi-illuminant (LSMI) dataset demonstrate the superiority of the proposed approach over state-of-the-art methods and commonly used loss functions when applied to the same architecture. Qualitative and quantitative evaluations highlight its effectiveness in achieving perceptually accurate WB correction, particularly in multi-illuminant environments. By bridging statistical modeling with modern deep learning, this work establishes the critical role of feature distribution alignment in advancing WB correction and sets a new benchmark for robustness and generalization in complex lighting scenarios.

Accurate and fast 3D object detection plays a role of paramount importance for safe and capable autonomous machines. LiDAR point cloud based methods have demonstrated impressive results, yet expensive LiDAR sensors make such approaches infeasible for wide-scale adaptation. Camera based methods on the other hand are performing sub-optimally given safety and accuracy requirements. Traditionally, camera based 3D object detection is performed by generating pseudo-LiDAR point clouds from RGB-D data and using point-cloud based models, however, irregular nature of point cloud data representation makes it challenging to exploit spatial local correlations on 3D space and point cloud based models generally suffer from this. To this end, we propose Sparse Voxel based 3D Object Detection, our approach differs from traditional approaches by converting point cloud information to sparse voxel grid and utilizing sub-manifold sparse convolutions to extract information instead of PointNet based models. Furthermore, we propose Bin-Mixing layers. Bin-Mixing replaces the output layer of a neural network and boosts performance by representing the problem of regression in a fashion that is easier for network to learn.