Accurate 3D hand pose and shape estimation from monocular RGB images is crucial for egocentric vision applications such as augmented reality, human-computer interaction, and robotics. This task is very challenging due to frequent hand occlusions and cluttered, dynamic backgrounds typical in first-person views. To address these issues, we introduce a unified RGB-only framework that eliminates the need for depth sensors or multi-view setups. Our approach comprises three key components: (1) a transformer-based SHARP module that generates pseudo-depth maps for hand segmentation, (2) a 3D hand mesh reconstruction pipeline based on the MANO model with context-aware refinement for occlusion handling, and (3) a spatiotemporal transformer that integrates 3D hand poses with 2D object cues for robust action recognition. We evaluated the proposed model on the H2O dataset, achieving 96.53
Deep Learning Algorithms are becoming predominant in Remote sensing, particularly in Hyperspectral Images (HSI), while Convolutional Neural Networks (CNN) and their variations are widely employed due to their proficiency in extracting local features, their limitations in capturing sequential properties of spectral characteristics have led to increased attention on transformers. In computer vision, there’s a notable shift towards Vision Transformer (ViT)-based models. However, these models lack image-specific inductive bias, such as translational equivariance and locality. The application of transformers in HSI faces a critical challenge due to the absence of suitable pre-processing and optimization methods, significantly impacting the overall performance of the models. To address this, we adopt a ViT-based backbone and incorporate pre-processing strategies inspired by spectral–spatial residual networks (SSRN) to effectively extract and analyze the spectral and spatial features of hyperspectral data. By combining these techniques, the proposed model demonstrates a significant improvement in accuracy, even when trained on fewer samples compared to previous approaches. The implementation code is available at github.com/ShashNagendra/HSI-Image-Classification-using-CNN-and-ViT .
Epileptic seizure detection using EEG is critical for neurological diagnosis and management. While deep learning methods have shown promise, they often fail to capture the temporal dependencies within EEG signals. This study proposes a Graph Neural Network (GNN)-based approach tailored for single-channel EEG classification using the University of Bonn (BONN) dataset. Each EEG signal is divided into eight non-overlapping segments, from which features are extracted. These segments are represented as graph nodes, with edges formed based on feature similarity between segments. The resulting graph representation enables classification into five distinct EEG categories (Z, N, F, S, O). To ensure robust evaluation, we trained and validated the proposed Graph Convolutional Network (GCN) model using 10 different random seeds. The best performance was achieved with seed 4, yielding an accuracy of 91.00
Large Language Models (LLMs) are widely used for tasks like translation and conversation but often reflect cultural biases from their training data. We introduce two frameworks, “The Life of X” and “RECAP”—to evaluate how LLMs handle culturally grounded prompts. The Life of X explores how identity cues (e.g., names) influence narrative framing, while RECAP tests model sensitivity to prompt specificity in recommendation tasks using semantic and manual metrics. Evaluating models like GPT-4, Gemini, and Claude, we find consistent identity-driven variation and a tendency to default to Western norms unless explicitly guided. These findings highlight the need for culturally aware evaluation methods in LLM development.
Segmenting plant root systems is vital in high-throughput phenotyping, providing insights into morphology, growth patterns, and stress responses. However, generating dense pixel-wise annotations for training segmentation models is labor-intensive and often impractical, especially for large datasets. To address this challenge, we propose a weakly supervised framework for root segmentation that relies only on sparse point-level annotations. Our model is based on a UNet architecture with a ResNet encoder and incorporates a composite loss function that combines binary cross-entropy, Dice loss, and consistency loss. This enables the model to learn accurate root structures even under limited supervision. We evaluate the proposed method on the ChronoRoot dataset and compare it with standard segmentation baselines. Despite relying on limited annotations, our method achieves competitive or superior performance to fully supervised and topology-aware models in Dice and precision scores. Through detailed ablation studies, we also demonstrate the individual contributions of each loss component to the overall performance. The proposed approach reduces annotation efforts while maintaining high segmentation accuracy, making it suitable for large-scale plant phenotyping applications.
