Hojun Lee

Co-founder & CEO at Xperty Corp.

hojun.lee @ xperty.co.kr

Scholar  /  LinkedIn

Bio

I received Ph.D. in 2024 at Department of Intelligence and Information from Seoul National University under the supervision of Prof. Nojun Kwak.

Research

My research centers on visual recognition, with a strong focus on object detection. I have explored few-shot object detection to enable robust recognition from limited data, and proposed principled approaches for coreset selection in complex, multi-object scenarios. Beyond detection, I’ve worked on safeguarding generative models by developing targeted data protection techniques for diffusion models—offering fine-grained control and traceability against unauthorized data use.

We address the lack of controllable and traceable protection in diffusion models by aligning fine-tuning trajectories with adversarially perturbed data, enabling intentional identity or pattern-level transformations in the generated outputs.

We present ARC-NeRF, leveraging Area Ray casting to cover broader unseen views with a single ray and adaptive high-frequency regularization. Additionally, luminance consistency regularization uses relative luminance as 'free lunch' data to improve texture accuracy.

We tackle the challenge of coreset selection in multi-object images by introducing a submodular selection strategy that captures both representativeness and diversity across classes.

We address the heuristic nature of end-to-end detection training by reformulating it as a density estimation task, using a mixture model with probabilistic loss functions to improve reliability.

We aim to improve few-shot detection under high intra-class variance by leveraging per-sample prototypes and attention-based refinement to preserve distinctive support information.

We avoid the complexity of anchor-based matching by modeling bounding box distributions directly, enabling a simpler and more principled training process for object detection.

We address the difficulty of predicting hypotension during dialysis by training a recurrent model on large-scale timestamped vitals to capture complex temporal patterns in real time.

We revisit autoencoder-based generation by learning the latent distribution explicitly, removing the need for a predefined prior and improving sample quality through density estimation.