3 papers indexed
Kwonyoung Ryu, In-Jae Lee, Jonghyun Jin, Hyunjee Lee, Jongmin Lee, Jaesik Park
Large view synthesis models synthesize novel views through cross-view attention without explicit 3D representations, and recent studies have shown that they learn accurate spatial correspondence from RGB supervision alone. We observe that this correspondence generalizes beyond ap…
Minhyuk Hwang, Sangmin Kim, Seunguk Do, Daneul Kim, Jaesik Park
Existing volumetric capture of dynamic human performance achieves high fidelity with dense camera arrays. However, in real-world scenarios, only a handful of low-overlap cameras are available, which degrades the output quality and leaves large areas unobserved. Recent 4D reconstr…
Eunsung Cha, Hyunjoon Lee, Jaesik Park
Open-vocabulary 3D Gaussian segmentation is challenging because it requires language understanding for diverse queries and accurate separation of Gaussians along object boundaries. Prior approaches either embed language knowledge into individual Gaussians to improve query respons…