Sunghwan Kim

I am a first-year PhD student at UC San Diego. Previously, I was a First Lieutenant at the Agency for Defense Development (ADD), the South Korean counterpart to the U.S. DARPA. I received my B.S. in Electrical Engineering and Mathematics (double major) at KAIST.

Email  /  CV  /  Linkedin  /  Google Scholar  /  Github

In this paper, we demonstrate that the text embeddings of vision-language models encode domain-invariant semantic knowledge at the pixel-level and propose a framework to utilize this information.

In this paper, we emphasize the significance of leveraging texture information to enhance performance in domain generalized semantic segmentation.

In this paper, we designed data gathering trials for infrared images from air-to-ground military systems, which are used for AI training.

We developed multi-threading C++ software that optimizes CPU and NPU resources during the inference phase of ML models, interfacing with the flight control unit of UAVs. Also, we devised a test-time diversification method for time-series images to reduce the predictive uncertainty of the models.

We implemented model compression techniques such as feature distillation and structural pruning to accelerate ML models on edge devices, including NPU, FPGA board and edge GPU. We were able to double the throughput of edge-devices by applying the techniques.

We designed real-time object detection algorithms for UAVs. We generated synthetic infrared images using a 3D engine for training data, and established an end-to-end training pipeline.

webpage template from Jon Barron