Hello! I'm Clayton Bromley, an experienced deep learning and computer vision engineer with a strong foundation in building AI-driven solutions that bridge technology and real-world applications. My background in entrepreneurship fuels my passion for developing scalable, impactful solutions, enabling me to approach challenges from both technical and business perspectives.
I specialize in designing, training, and deploying machine learning models that solve complex visual understanding problems, from object detection to video segmentation, with a focus on performance and accuracy. By combining my technical expertise with a strategic mindset, I aim to drive innovation and create meaningful advancements in the fields of AI and computer vision.
Publications
An Analysis of Segment Anything 2
Video object segmentation (VOS) is essential for video perception, with Meta AI's Segment-Anything Model 2 (SAM 2) leading as the state-of-the-art architecture. SAM 2 excels at segmenting objects in both clean and transformed video data by progressively filtering noise and emphasizing objects of interest. This study contributes by creating complex transformation datasets, analyzing SAM 2's interpretation at each stage, and visualizing object segmentation to improve real-world VOS performance in cluttered and obscured scenes.
Addressing Issues with Working Memory in Video Object Segmentation
This study introduces a simple algorithmic adjustment to enhance video object segmentation (VOS) models by improving their handling of inconsistent camera views, such as sudden cuts or extreme context shifts, which typically disrupt working memory in VOS. By selectively updating memory only when the target object is in frame, this modification significantly enhances model performance on challenging video data, thereby broadening the practical applicability of VOS models in real-world scenarios.
Presentations
ClayNet: Improved VOS Modeling
Video object segmentation (VOS) is a challenging task in the world of deep learning and computer vision and is essential for the proper perception and analysis of videos. Cutie, the current state-of-the-art model for VOS, performs very well on clean video. When artificially introducing long-term obscurations of the segmented object, Cutie’s performance drops significantly. This project aims to design custom DAVIS datasets with artificial obscurations and propose a ClayNet model with improved performance.
Papers
Passive Sonar Source Classification and Localization
This paper proposes a model pipeline to enhance passive sonar navigation for submarines and naval vessels, addressing challenges in contact classification, localization, and distance approximation under noisy aquatic conditions. By combining convolutional neural networks (CNN) for signal discrimination, 2D image filtering for bearing determination, and a novel LSTM model for distance estimation based on bearing rate alone, this approach aims to advance towards fully autonomous sonar-based navigation.
Deepfake Modeling
This paper explores different models for generating Deepfake images while also exploring how facial features can be manipulated to create convincing fake images. Through implementing two different models for facial generation as well as two different forms of facial manipulation, the paper compares and contrasts the different techniques used to generate fake faces through the use of GANs, diffusion models, and encoder-decoder architectures.
Projects
Towster Website
Towster, a defense medical technology startup, is pioneering casualty evacuation innovations for challenging denied-airspace combat scenarios, specializing in maritime and amphibious operations. The Towster pod offers en route hyperbaric oxygen therapy and covert underwater evacuation capabilities, enhancing survival rates and ensuring patient safety in hostile environments. As a cofounder, this project entails the creation of the company website and the design of the project and associated research.
