The first trajectory-level ergodic formulation for active Gaussian scene reconstruction. Footprint-aware ergodic search gains +1.5 dB PSNR over baselines and runs directly on real robots, including the Unitree GO2 and Franka FR3.

High-rate event camera simulation accelerated by motion-vector frame interpolation, with HDR support, sensor noise models, motion blur, and tools for downstream training.

Zero-shot wide-baseline feature matching for event cameras, using motion-robust multi-timescale attention and sparsity-aware token selection. Improves on prior event matching methods by 37.7%.

Human motion as a continuous-time field from events. A recurrent, time-implicit network supports parallel pose queries at any temporal resolution, with 23.8% lower joint error and 69% less compute.

Neural inertial odometry that exploits roto-reflective symmetries in IMU data through equivariant, gravity-aligned canonicalization, improving displacement estimates on the TLIO, Aria, RIDI, and OxIOD benchmarks.

Continuous color video reconstructed from a single RGB frame and an event stream, combining long-range motion modeling with neural synthesis to outperform baselines by 3.61 dB PSNR.

Unsupervised event-based independent motion segmentation that uses geometric constraints to generate pseudo-labels, handles any number of moving objects, and is competitive with supervised methods on EVIMO.

Self-supervised dense continuous-time motion estimation for event cameras, combining contrast maximization with non-linear motion priors to improve zero-shot real-world performance by 29% on EVIMO2.

Global trajectory and body motion of 3D humans from in-the-wild videos, combining robustified SLAM for camera motion with a video transformer for kinematic body regression.

Test-time optimization for dense video point tracking with an invertible deformation network and foundation-model priors, over 10× faster and more robust than OmniMotion.

An event camera dataset spanning ground vehicles, legged robots, and aerial platforms in challenging conditions, with synchronized stereo event cameras, LiDAR, IMU, and RTK.

3D reconstruction from event cameras via continuous visual hulls built from apparent contour events, producing high-fidelity meshes from real event sequences with per-event refinement.

High-speed object catching with low-latency event-based neural networks and a binary event history image representation, reaching an 81% catch rate on balls at up to 13 m/s.

GAN-based event simulation from image pairs with cycle-consistency losses, enabling object detection and pose estimation networks to train on large image datasets without real event labels.

A higher-order function network that generates cost-to-go functions over configuration space, running orders of magnitude faster than planners that explore configuration space online.

Single-image surface reconstruction with a higher-order function network that maps unit-sphere samples to local tangent planes, generating surfaces at arbitrary resolution with few parameters.

Single-view 3D reconstruction that lifts points into an embedding where Euclidean distance equals geodesic distance on the surface, improving surface normals and enabling unsupervised object decomposition.

A temporal normalization transform for event data that makes spatiotemporal convolutions equivariant to optical flow, removing the need for motion augmentation in classification.

A hybrid unsupervised egomotion pipeline combining learned optical flow and disparity with RANSAC outlier rejection and least-squares pose estimation, with automatic moving-object segmentation on KITTI.