Antonio Robles-Kelly

Despite significant advancements in computer vision, semantic segmentation models may be susceptible to backdoor attacks. These attacks, involving hidden triggers, aim to cause the models to misclassify instances of the victim class as the target class when triggers are present, posing serious threats to the reliability of these models. To further explore the field of backdoor attacks against semantic segmentation, in this paper, we propose a simple yet effective backdoor attack called Contextual Segmentation Backdoor Attack (ConSeg). ConSeg leverages the contextual information inherent in semantic segmentation models to enhance backdoor performance. Our method is motivated by an intriguing observation, i.e., when the target class is set as the `co-occurring' class of the victim class, the victim class can be more easily `mis-segmented'. Building upon this insight, ConSeg mimics the contextual information of the target class and rebuilds it in the victim region to establish the contextual relationship between the target class and the victim class, making the attack easier. Our experiments reveal that ConSeg achieves improvements in Attack Success Rate (ASR) with increases of 15.55\%, compared to existing methods, while exhibiting resilience against state-of-the-art backdoor defenses.

For the sentence-level sentiment classification task, learning Contrastive Discourse Relations (CDRs) like a-but-b is difficult for Deep Neural Networks (DNNs) via purely data-driven training. Several methods exist in the literature for dissemination of CDR information with DNNs, but there is no dedicated dataset available to effectively test their dissemination performance. In this paper, we propose a new large-scale dataset for this purpose called Covid19-twitter, which contains around 100k tweets symmetrically divided into various categories. Instead of manual annotation, we used a combination of an Emoji analysis and a lexicon-based tool called Valence Aware Dictionary and sEntiment Reasoner (VADER) to perform automatic labelling of the tweets, while also ensuring high accuracy of the annotation process through some quality checks. We also provide benchmark performances of several baselines on our dataset for both the sentiment classification and CDR dissemination tasks. We believe that this dataset will be valuable for discourse analysis research in sentiment classification.

Normal estimation on 3D point clouds is a fundamental problem in 3D vision and graphics. Current methods often show limited accuracy in predicting normals at sharp features (e.g., edges and corners) and less robustness to noise. In this paper, we propose a novel normal estimation method for point clouds. It consists of two phases: (a) feature encoding which learns representations of local patches, and (b) normal estimation that takes the learned representation as input and regresses the normal vector. We are motivated that local patches on isotropic and anisotropic surfaces have similar or distinct normals, and that separable features or representations can be learned to facilitate normal estimation. To realise this, we first construct triplets of local patches on 3D point cloud data, and design a triplet network with a triplet loss for feature encoding. We then design a simple network with several MLPs and a loss function to regress the normal vector. Despite having a smaller network size compared to most other methods, experimental results show that our method preserves sharp features and achieves better normal estimation results on CAD-like shapes.

Deep ensemble models have garnered attention for their capacity to yield robust predictions along with well-calibrated uncertainty estimates. Empirical evidence suggests that augmenting the number of models within deep ensembles fosters greater diversity in the functional space, enabling the exploration of multiple local optima effectively. This leads to more robust uncertainty estimates in the model. In this paper, we introduce a novel methodology to substantially enhance model diversity by inducing model heterogeneity through controlled perturbation of input images, thereby de-correlating individual models based on their local bias. Drawing inspiration from the bias-variance trade-off, our approach incorporates pre-generated masks during the training process to further refine model performance. To validate the efficacy of our technique in providing more robust estimates under dataset shift, we evaluate model diversity and predictive uncertainty on test samples from both familiar and unfamiliar distributions. Our results demonstrate that our proposed method outperforms recent approaches in detecting out-of-distribution samples, establishing its superiority within the existing literature.

Hierarchical image classification is an arduous task in deep learning and computer vision. It requires classifying multiple image classes following a taxonomy or data hierarchy. This paper introduces a bottom-up hierarchical capsule network (BUH-CapsNet) designed to address hierarchical multi-label classification. The hierarchical structure of BUH-CapsNet allows it to build a tree-like structure for classification problems, making use of the data hierarchy. This structure enables the network to learn more complex relationships in the taxonomy by balancing the hierarchical levels and following the fine-to-coarse paradigm, leading to more accurate classification results. Furthermore, the bottom-up architecture of the BUH-CapsNet enforces hierarchical consistency, using the hierarchical structure of the datasets. We trained our BUH-CapsNet considering the hierarchical level weights that keep a balance between the levels. Experiments on six widely available datasets show that BUH-CapsNet achieves better results than existing multi-label classification methods and performs better when handling hierarchical labels.

Squamous cell carcinoma is the most common type of cancer that occurs in squamous cells of epithelial tissue. Histopathological evaluation of tissue samples is the gold standard approach used for carcinoma diagnosis. SCC detection based on various histopathological features often employs traditional machine learning approaches or pixel-based deep CNN models. This study aims to detect keratin pearl, the most prominent SCC feature, by implementing RetinaNet one-stage object detector. Further, we enhance the model performance by incorporating an attention module. The proposed method is more efficient in detection of small keratin pearls. This is the first work detecting keratin pearl resorting to the object detection technique to the extent of our knowledge. We conducted a comprehensive assessment of the model both quantitatively and qualitatively. The experimental results demonstrate that the proposed approach enhanced the mAP by about 4% compared to default RetinaNet model.