Lecture 27: Inverse Graphics/Research In Angjoo's Group (17)
han20192019
I want to confirm my understanding for this topic. Is it that we can recover the motion/different perspectives according to a single picture? I think it's very cool and the 3D reconstruction can be applied in many fields, like VR
philippe-eecs
Do you need a priors on the 3D model to construct this? Obviously we know a ton about the human shape, but this seems pretty much impossible to do without prior info on the subject you are constructing. You would usually need multiple images to properly construct a 3D model.
ronthalanki
@philippe-eecs I'm not sure about the specifics on the model being used here, but generally the state-of-the-art neural nets are somewhat able to generate 3d models of the human body without any prior knowledge of the specific human. Prof. Kanazawa has a more recent paper here that proposes an algorithm to do this: https://akanazawa.github.io/hmr/
I want to confirm my understanding for this topic. Is it that we can recover the motion/different perspectives according to a single picture? I think it's very cool and the 3D reconstruction can be applied in many fields, like VR
Do you need a priors on the 3D model to construct this? Obviously we know a ton about the human shape, but this seems pretty much impossible to do without prior info on the subject you are constructing. You would usually need multiple images to properly construct a 3D model.
@philippe-eecs I'm not sure about the specifics on the model being used here, but generally the state-of-the-art neural nets are somewhat able to generate 3d models of the human body without any prior knowledge of the specific human. Prof. Kanazawa has a more recent paper here that proposes an algorithm to do this: https://akanazawa.github.io/hmr/