Suhwan Cho

I am a Ph.D. student in the Image and Video Pattern Recognition Lab. at Yonsei University, Seoul, Korea. I was a Research Scientist Intern in the Deep Learning Group at Adobe Research, San Jose, California, USA. My research mainly focuses on video editing technologies such as video object segmentation and video inpainting.

I am always open to collaborations! Please feel free to contact me if you are interested :)

CV / GitHub / Linkedin / Google Scholar

We propose a dual-branch architecture that hierarchically incorporates both structural priors and appearance information.

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

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

We propose the inter-frame curve network to effectively leverage the spatio-temporal dependency of the human skeleton.

We propose a novel network architecture that operates regardless of motion availability, termed as a motion-as-option network. A collaborative network learning strategy is also introduced to fully exploit the property of the proposed network.

We propose a prototype memory network architecture to effectively extract and store appearance and motion information based on component prototypes from RGB images and optical flow maps.

We propose three novel strategies to suppress the negative influence of background distractions in semi-supervised video object segmentation.

We propose a prototype sampling network that only samples prototypes corresponding to obtain robustness to inconsistencies between RGB images and depth maps.

We propose an implicit two-path AE, a structure in which two encoders implicitly model appearance and motion features, along with a single decoder that combines them to learn normal video patterns.

We introduce a bijective matching mechanism to find the best matches from the query frame to the reference frame and also vice versa. A mask embedding module is also proposed to consider multiple historic masks simultaneously.