It achieves clearer object separation than Feature-3DGS.
Kosta Derpanis (sabbatical in Zurich)#932
(1/5) Excited to introduce MLP-Splatting, a novel way to decompose a scene into object-/part-level primitives!
Each object can be represented with a few tens of light-field-based primitives (c.f. 3DGS: hundreds+).
w/ @cheng_yuzh21715 @XinKong_IC @paulhjkelly @AjdDavison
Users express excitement about MLP-Splatting for object-level scene decomposition with light-field primitives, calling the technique sick and exciting.
MLP-Splatting, decomposed reconstruction and Splatting from our lab at Imperial College. Come and find Shinjeong at CVPR if you are interested.
(1/5) Excited to introduce MLP-Splatting, a novel way to decompose a scene into object-/part-level primitives!
Each object can be represented with a few tens of light-field-based primitives (c.f. 3DGS: hundreds+).
w/ @cheng_yuzh21715 @XinKong_IC @paulhjkelly @AjdDavison
(2/5) Project Page: http://shinjeongkim.com/mlp-splatting
Thanks to the expressive representation on local primitives, we achieve comparable NVS quality but with 3X rendering speed and with 1/15 memory.
(3/5) Having small number of meaningful primitives make object selection and scene editing easy! We can even select objects with a few clicks (video: <6 click per chess-piece).
(4/5) Or by language if guided with language feature during training!
(5/5) We can apply SE(3), deletion, and color change to per selected primitives.
@shinjeong99 @deepfry_n @cheng_yuzh21715 @XinKong_IC @paulhjkelly @AjdDavison dawg this is so sick. exciting stuff
@shinjeong99 @cheng_yuzh21715 @XinKong_IC @paulhjkelly @AjdDavison The “few meaningful primitives” part feels very important. A lot of 3DGS work focuses on visual quality, but object-level structure and editability may matter more if these scenes are going to become usable assets or interfaces.
It achieves clearer object separation than Feature-3DGS.
Kosta Derpanis (sabbatical in Zurich)#932(1/5) Excited to introduce MLP-Splatting, a novel way to decompose a scene into object-/part-level primitives!
Each object can be represented with a few tens of light-field-based primitives (c.f. 3DGS: hundreds+).
w/ @cheng_yuzh21715 @XinKong_IC @paulhjkelly @AjdDavison