Published in ICCV, 2023
We use continuous diffusion to generate 3d point clouds, both unconditionally and based on monocular images. We propose a novel scheme for image conditioning via projected 2d feature lookup and show benefits of continuous diffusion, such as efficient upsampling of point clouds and computing sample probabilities.
Download here
Published in ECCV, 2022
We use attention to pool information from multiple posed input views in a global voxel representation. This allows for 3d object detection, shape reconstruction, novel view synthesis and more. A key trick is to exploit the image formation geometry to restrict the attention matrix, making the approach scalable and robust.
Download here
Published in NeurIPS, 2020
Local features are difficult to learn because they are inherently discrete. DISK uses reinforcement learning to get stable gradients and learn the task end-to-end.
Download here
Published in Dalton Transactions, 2015
Application of a patented synthesis protocol to a number of borohydride compounds. My BSc thesis in chemistry revolved around synthesis of [Ph4P]2[Mn(BH4)4], discussed in this article.
Download here