Tag: #3d-reconstruction15 results
CALHippo utilizes a custom segmentation pipeline, including CellPoseSAM, to map human hippocampal cells in 3D. The project bridges high-resolution slice annotations with low-resolution density estimation to reconstruct cellular point clouds.
CVPR 2026 research highlights a shift in autonomous driving from simple perception to controllable scene generation and end-to-end decision-making. New methods like HorizonForge and DiffusionHarmonizer enable precise trajectory editing and photorealistic simulation enhancement.
CVPR 2026 papers shift computer vision from benchmark optimization to real-world adaptability, challenging frozen models, predefined goals, ample data, and structured inputs. Highlights include LIT for live interactive training in video segmentation, INSID3 for training-free in-context segmentation using DINOv3, and MegaDepth-X for sparse photo reconstruction.
RAM is a new unified framework for 3D human motion reconstruction in complex multi-person video scenes, tackling ID switches, tracking losses, and discontinuities. It integrates SegFollow for stable tracking, T-HMR for temporal reconstruction, motion prediction, and adaptive fusion. Achieves superior zero-shot performance on PoseTrack benchmark.
Embodied AI community open-sourced a SOTA method for infinite-frame video to real-time 3D world reconstruction. Robots now process endless footage on-the-fly like Naruto's Byakugan. Enables advanced perception for robotics.
Hangzhou Zhuma Innovation raised tens of millions in angel funding led by Fengrei Capital for consumer-grade 3D reconstruction hardware using 3DGS tech. First product Pebble targets pro designers with portable, affordable scanning. Plans expand to consumer 'space memory camera' and global markets.
Apple's AI research team released LiTo (Surface Light Field Tokenization), solving key 3D reconstruction challenges. It generates complete 3D objects from a single flat image while accurately restoring multi-view lighting effects. This eliminates the need for multi-angle image inputs in traditional methods.
Peking University and Gaode Maps jointly developed Orbit2Ground, a generative photogrammetry method that reconstructs high-fidelity 3D city models from sparse satellite images. It leverages urban geometric priors and generative AI to overcome extreme viewpoint extrapolation challenges from top-down to ground-level perspectives. This enables efficient creation of digital city twins for gaming, drone logistics, and emergency systems.
This article explores how to use NVIDIA Nsight developer tools to optimize the Omniverse NuRec pipeline. It focuses on improving the performance of 3D environment reconstruction from multisensor data like lidar and cameras.
Room360 is a new platform designed for transforming standard video input into 3D spatial reconstructions. It leverages advanced computer vision to enable immersive environment modeling.