A5013897558- Advanced Vision and Imaging
- Advanced Neural Network Applications
- Computer Graphics and Visualization Techniques
- Advanced Image Processing Techniques
- Model Reduction and Neural Networks
- Domain Adaptation and Few-Shot Learning
- Generative Adversarial Networks and Image Synthesis
- Human Pose and Action Recognition
- 3D Shape Modeling and Analysis
- Image and Signal Denoising Methods
- Neural Networks and Applications
- Anomaly Detection Techniques and Applications
- Multimodal Machine Learning Applications
- Advanced Image and Video Retrieval Techniques
- Lattice Boltzmann Simulation Studies
- Robotics and Sensor-Based Localization
- Medical Image Segmentation Techniques
- Industrial Vision Systems and Defect Detection
- Video Surveillance and Tracking Methods
- Autonomous Vehicle Technology and Safety
- CCD and CMOS Imaging Sensors
- Image Enhancement Techniques
- Gaussian Processes and Bayesian Inference
- Natural Language Processing Techniques
- Fire Detection and Safety Systems
Yonsei University
2025
Sungkyunkwan University
2022-2024
Korea Advanced Institute of Science and Technology
2024
University of North Carolina at Chapel Hill
2015-2020
Adobe Systems (United States)
2019
University of North Carolina Health Care
2016
We present a transformation-grounded image generation network for novel 3D view synthesis from single image. Our approach first explicitly infers the parts of geometry visible both in input and views then casts remaining problem as completion. Specifically, we predict flow to move pixels along with visibility map that helps deal occulsion/disocculsion. Next, conditioned on those intermediate results, hallucinate (infer) object invisible In addition new structure, training combination...
We present a new public dataset with focus on simulating robotic vision tasks in everyday indoor environments using real imagery. The includes 20,000+ RGB-D images and 50,000+ 2D bounding boxes of object instances densely captured 9 unique scenes. train fast category detector for instance detection our data. Using the we show that, although increasingly accurate fast, state art is still severely impacted by scale, occlusion, viewing direction all which matter robotics applications. next...
Although deep convolutional neural networks (CNNs) have shown remarkable results for feature learning and prediction tasks, many recent studies demonstrated improved performance by incorporating additional handcrafted features or fusing predictions from multiple CNNs. Usually, these combinations are implemented via concatenation averaging output scores several In this paper, we present new approaches combining different sources of knowledge in learning. First, propose amplification, where...
In this paper, we introduce a new dataset consisting of 360,001 focused natural language descriptions for 10,738 images. This dataset, the Visual Madlibs is collected using automatically produced fill-in-the-blank templates designed to gather targeted about: people and objects, their appearances, activities, interactions, as well inferences about general scene or its broader context. We provide several analyses demonstrate applicability two description generation tasks: generation,...
Autonomous vehicles are an exemplar for forward-looking safety-critical real-time systems where significant computing capacity must be provided within strict size, weight, and power (SWaP) limits. A promising way forward in meeting these needs is to leverage multicore platforms augmented with graphics processing units (GPUs) as accelerators. Such approach being strongly advocated by NVIDIA, whose Jetson TX1 board currently a leading multicore+GPU solution marketed autonomous systems....
Neural radiance fields (NeRF) have demonstrated the potential of coordinate-based neural representation (neural or implicit representation) in rendering. However, using a multi-layer perceptron (MLP) to represent 3D scene object requires enormous computational resources and time. There been recent studies on how reduce these inefficiencies by additional data structures, such as grids trees. Despite promising performance, explicit structure necessitates substantial amount memory. In this...
In this paper, we introduce a new dataset consisting of 360,001 focused natural language descriptions for 10,738 images. This dataset, the Visual Madlibs is collected using automatically produced fill-in-the-blank templates designed to gather targeted about: people and objects, their appearances, activities, interactions, as well inferences about general scene or its broader context. We provide several analyses demonstrate applicability two description generation tasks: generation,...
Computer vision has achieved impressive progress in recent years. Meanwhile, mobile phones have become the primary computing platforms for millions of people. In addition to phones, many autonomous systems rely on visual data making decisions, and some these limited energy (such as unmanned aerial vehicles also called drones robots). These batteries, efficiency is critical. This paper serves following two main purposes. First, examine state art low-power solutions detect objects images....
Neural fields, also known as coordinate-based or implicit neural representations, have shown a remarkable capability of representing, generating, and manipulating various forms signals. For video however, mapping pixel-wise coordinates to RGB colors has relatively low compression performance slow convergence inference speed. Frame-wise representation, which maps temporal coordinate its entire frame, recently emerged an alternative method represent videos, improving rates encoding While...
Physics-informed neural networks (PINNs) have recently emerged as promising data-driven PDE solvers showing encouraging results on various PDEs. However, there is a fundamental limitation of training PINNs to solve multi-dimensional PDEs and approximate highly complex solution functions. The number points (collocation points) required these challenging grows substantially, but it severely limited due the expensive computational costs heavy memory overhead. To overcome this issue, we propose...
Generative Artificial Intelligence (AI) has rapidly advanced the field of computer vision by enabling machines to create and interpret visual data with unprecedented sophistication. This transformation builds upon a foundation generative models produce realistic images, videos, 3D or 4D content. Traditionally, primarily focus on fidelity while often neglecting physical plausibility generated gap limits their effectiveness in applications requiring adherence real-world laws, such as robotics,...
Large-scale generative models, such as text-to-image diffusion have garnered widespread attention across diverse domains due to their creative and high-fidelity image generation. Nonetheless, existing large-scale models are confined generating images of up 1K resolution, which is far from meeting the demands contemporary commercial applications. Directly sampling higher-resolution often yields results marred by artifacts object repetition distorted shapes. Addressing aforementioned issues...
3D super-resolution aims to reconstruct high-fidelity models from low-resolution (LR) multi-view images. Early studies primarily focused on single-image (SISR) upsample LR images into high-resolution However, these methods often lack view consistency because they operate independently each image. Although various post-processing techniques have been extensively explored mitigate inconsistencies, yet fully resolve the issues. In this paper, we perform a comprehensive study of by leveraging...
Continuous convolution has recently gained prominence due to its ability handle irregularly sampled data and model long-term dependency. Also, the promising experimental results of using large convolutional kernels have catalyzed development continuous since they can construct very efficiently. Lever-aging neural networks, more specifically multilayer perceptrons (MLPs), is by far most prevalent approach implementing convolution. However, there are a few drawbacks, such as high computational...
3D Gaussian splatting (3DGS) has recently emerged as an alternative representation that leverages a Gaussian-based and introduces approximated volumetric rendering, achieving very fast rendering speed promising image quality. Furthermore, subsequent studies have successfully extended 3DGS to dynamic scenes, demonstrating its wide range of applications. However, significant drawback arises following methods entail substantial number Gaussians maintain the high fidelity rendered images, which...
We investigate using reinforcement learning agents as generative models of images (extending arXiv:1804.01118). A agent controls a simulated painting environment, and is trained with rewards provided by discriminator network simultaneously to assess the realism agent's samples, either unconditional or reconstructions. Compared prior work, we make number improvements architectures discriminators that lead intriguing at times surprising results. find when sufficiently constrained, can learn...
Recent studies in Radiance Fields have paved the robust way for novel view synthesis with their photorealistic rendering quality. Nevertheless, they usually employ neural networks and volumetric rendering, which are costly to train impede broad use various real-time applications due lengthy time. Lately 3D Gaussians splatting-based approach has been proposed model scene, it achieves remarkable visual quality while images real-time. However, suffers from severe degradation if training blurry....