News
Publications
No publications match these filters.
SSC-Priors: Exploring Semantic and Visibility Priors to Boost Lidar Semantic Scene Completion
The journal extension of SSC-Priors shows that off-the-shelf semantic pseudo-labels and ray-cast visibility, fed only as extra inputs, consistently boost four lidar SSC networks on SemanticKITTI and SSCBench-nuScenes, reaching state-of-the-art mIoU among reproducible single-frame methods. A model trained once on ground-truth labels works with any segmenter, and oracle experiments show semantics is still the main bottleneck.
Exploring Easy Boosts for Lidar Semantic Scene Completion
This paper, a.k.a. SSC-Priors, is a study of plug-and-play semantic and visibility input priors that boost four lidar SSC baselines by +5.2 mIoU on average on SemanticKITTI. With no architectural redesign beyond the first layer needed, we reach SOTA in mIoU among reproducible single-frame methods.
LiDPM: Rethinking Point Diffusion for Lidar Scene Completion
LiDPM is a diffusion model for scene-level lidar completion that leverages a vanilla DDPM avoiding the need for local diffusion approximations.
Information compression method and apparatus
This method introduces a control circuit and neural network pipeline that compresses 3D source field information into compact vector representations over subdivided spaces, enabling efficient field value estimation via a trained encoder-decoder model.
A Controlled Hahn-Mazurkiewicz Theorem and its Applications
This paper contains the proof of the controlled version of the classical Hahn-Mazurkiewicz Theorem that provides upper and lower bound of the Hölder dimension of a metric Peano continuum as a function of its S-dimension.
FEAR: Fast, Efficient, Accurate and Robust Visual Tracker
FEAR is a family of Siamese visual trackers that combine dual-template adaptation, pixel-wise fusion, and efficient backbones to achieve state-of-the-art accuracy and speed, while also introducing a new benchmark for evaluating energy-efficient tracking.
DAD-3DHeads: A Large-scale Dense, Accurate and Diverse Dataset for 3D Head Alignment from a Single Image
DAD-3DHeads is a large-scale, diverse dataset with over 3.5K 3D head landmarks, and DAD-3DNet is a robust model trained on it for dense 3D head alignment, achieving state-of-the-art results in pose, shape reconstruction, and landmark estimation across multiple benchmarks.
DeblurGAN-v2: Deblurring (Orders-of-Magnitude) Faster and Better
DeblurGAN-v2 is an end-to-end GAN for single image motion deblurring that combines a relativistic conditional GAN, double-scale discriminator, and Feature Pyramid Network to deliver state-of-the-art quality and efficiency across various backbones, enabling real-time performance and broader applicability to image restoration.
Kuratowski Monoids of n-Topological Spaces
The main result of this paper generalizes the famous Kuratowski 14-set closure-complement Theorem to polytopological spaces.