Search papers, labs, and topics across Lattice.
3
1
4
7
Task-structured routing in MoTE allows for interpretable and efficient multi-task video-language learning, outperforming traditional dense activation methods.
Achieving real-time 3D hand pose estimation without the need for camera parameters could revolutionize applications in AR and robotics.
Reconstructing realistic hand-object interactions from video just got an order of magnitude faster, thanks to a novel Gaussian Splatting approach that ensures physical consistency.
