Kiana Hooshanfar

Audio-visual saliency prediction aims to mimic human visual attention by identifying salient regions in videos through the integration of both visual and auditory information. Although visual-only approaches have significantly advanced, effectively incorporating auditory cues remains challenging due to complex spatio-temporal interactions and high computational demands. To address these challenges, we propose Dynamic Token Fusion Saliency (DTFSal), a novel audio-visual saliency prediction framework designed to balance accuracy with computational efficiency. Our approach features a multi-scale visual encoder equipped with two novel modules: the Learnable Token Enhancement Block (LTEB), which adaptively weights tokens to emphasize crucial saliency cues, and the Dynamic Learnable Token Fusion Block (DLTFB), which employs a shifting operation to reorganize and merge features. An audio branch processes raw audio signals, and both modalities are integrated using our Adaptive Multimodal Fusion Block (AMFB). Extensive evaluations on six audio-visual benchmarks demonstrate that DTFSal achieves state-of-the-art performance while maintaining computational efficiency.

The visibility of brand logos on packaging plays a crucial role in shaping consumer perception, directly influencing the product's success. Analyzing eye-tracking data across large groups of individuals is both costly and time-intensive. Therefore, there is a growing need to develop models that capture human visual attention behavior effectively. This paper introduces a framework that models attention in the human visual system to brand logos on packaging designs, to measure brand logo visibility and its impact on consumer perception. The proposed method consists of three main steps: logo detection with YOLOv8, a novel saliency prediction model tailored for the packaging context, and integration of both to provide a brand attention score. Comparing logo detection and saliency map prediction with state-of-the-art models shows the superiority of the proposed methods, and the brand attention score aligns with previous psychophysical studies on brand visibility.

In the domain of Natural Language Processing (NLP), the integration of Large Language Models (LLMs) with Retrieval-Augmented Generation (RAG) represents a significant advancement towards enhancing the depth and relevance of model-generated responses. This paper introduces a novel hybrid RAG framework that synergizes the Sentence-Window and Parent-Child methodologies with an innovative re-ranking mechanism, aimed at optimizing the query response capabilities of LLMs. By leveraging external knowledge sources more effectively, the proposed method enriches LLM outputs with greater accuracy, relevance, and information fidelity. Evaluation against benchmark datasets demonstrates superior performance over existing state-of-the-art RAG techniques, with responses that are contextually appropriate and faithful to the source material.

Eye-tracking technology has rapidly gained popularity as a revolutionary solution for individuals with severe physical disabilities, empowering them to engage in their daily activities with newfound independence and efficiency. By utilizing advanced eye-tracking systems, individuals with limited mobility are able to control various devices and interfaces simply by moving their eyes. This paper utilizes deep learning techniques to create a low-cost real-time eye-tracking interface for controlling systems. A smart wheelchair and a robotic arm have been developed to design an eye-tracker, aiming to address the challenges faced by paralyzed people with severe physical limitations. The results demonstrate that eye-tracking is both fast and accurate, making it an effective tool for improving interactions and accessibility for disabled individuals.