Dogyoon Lee

I am Ph.D candidate at Yonsei University in Seoul, where I work on computer vision and machine learning.

My research mainly focuses on various 3D computer vision tasks using images and point cloud. Recent research field is related to neural rendering and its applications.

I'm always open to collaborations or suggestions. Please feel free to contact me if you have any questions or suggestions. :)

Email  /  CV  /  Google Scholar  /  Github

We propose a guided slot attention network to reinforce spatial structural information and obtain better foreground–background separation.

We propose two novel prototype-based attention mechanisms to incorporate different modalities and frames via dense propagation across them.

We impose the physical constraints on the blurring kernel of neural radiance field to construct clean neural radiance field from blurry images.

We propose a hierarchically decomposed graph convolution with a novel hierarchically decomposed graph, which consider the sematic correlation between the joints and the edges of the human skeleton.

We propose a novel framework for unsupervised video object segmentation, which can utilize both contextual and motion information from adjacent frames, introducing a temporal and scale alignment network (TSANet).

We propose a novel framework for point cloud processing, which can utilize both local and global information from point clouds, introducing a multidimensional feature representation (Mkconv).

We propose rescaling module for the image compression network which is enhanced version of existing GDN with higher degree of freedom.

We propose a novel self-attention module called the Expanded Self Attention for lane detection in autonomous vehicle to be robust in challenging situations.

We propose a novel data augmentation method called Rigid Subset Mix (RSMix) which generates virtual mixed samples by replacing part of the sample with shape-preserved subsets from another sample.

We propose a novel post-processing method to remove false positives in 3D vehicle detection utilizing the characteristics of the LiDAR censor itself.

We propose novel bi-directional token-masking autoencoder (BTMAE) for referring image segmentation (RIS) to effectively utilize contextual information between language and visual features.

We propose enhanced deblurred neural radiance fields from sparse view settings for more practical applications considering real-world scenarios.

We propose novel cost volume refinement model for self-supervised multi-frame monocular depth estimation model.

We propose novel blur kernel for motion estimation based on neural ordinary differential function to construct the deblurred neural radiance fields.

This website's source code is borrowed from jonbarron's website.

Last updated Mar 2024.