Jinjie Mai is an M.S./Ph.D. student at KAUST, fortunately under the supervision of Prof. Bernard Ghanem. He has a broad research interest in 3D vision including topics like localization, reconstruction, and generation. He has published many papers at top-tier AI conferences such as CVPR/ICCV/ECCV/NeurIPS/ICLR. He has collectively earned over 1,000 citations. His current research now focuses on 4D reconstruction and generation.
KAUST
MSc Computer Science
Sun Yat-Sen University
BEng Computer Science and Technology
We repurpose video diffusion models to reconstruct 4D geometry (4D pointmaps) from monocular video.
We introduce global bundle adjustment for NeRFs, significantly improving the quality of novel view synthesis from sparse views with noisy poses or even without poses.
We propose multiview video diffusion models that generate high-quality multi-view videos for dynamic objects from text.
We propose to use both 2D and 3D diffusion models through SDS optimization to generate high-quality, textured 3D meshes from a single image in the wild.
We propose EgoLoc, a novel approach for localizing 3D objects from egocentric videos, achieving SOTA performance on the Ego4D VQ3D task with 87.12% success rate over previous SOTA of only 8%.
We revisit PointNet++ and propose PointNeXt, a next-generation PointNet architecture.
We propose EIL, a novel adversarial erasing technique for weakly supervised object localization.