Publications

KnobGen introduces a novel dual-pathway framework for sketch-based image generation, designed to adapt to varying levels of user skill and sketch complexity. By leveraging a Coarse-Grained Controller for high-level semantic understanding and a Fine-Grained Controller for detailed refinement, KnobGen ensures precise yet flexible image generation across a broad spectrum of sketch inputs. A key innovation in the framework is the inference-time knob, which allows users to dynamically control the level of adherence to the input sketch. This approach democratizes image generation by effectively balancing fine-grained precision and broader spatial structure, making it suitable for both amateur and professional use.
Sept, 2024.
PDF CODE

Segformer3D is a light-weight and efficient hierarchical Transformer designed for 3D volumetric segmentation. It calculates attention across multiscale volumetric features, and avoids complex decoders. Instead it uses a simple yet effective all-MLP decoder to aggregate local and global attention features to produce highly accurate segmentation masks.
CVPRW, 2024.
PDF CODE

This paper presents a probabilistic-based drift correction module for improving the accuracy of Visual Inertial Simultaneous Localization and Mapping (SLAM) systems. The proposed method treats positioning measurements as stochastic variables in a multivariate distribution, incorporating geospatial priors to correct accumulated drift error. By optimizing the mode of this distribution, the approach reduces drift by up to ten times during long traverses. The module is designed to be compatible with various SLAM systems, enhancing their performance in GPS-denied or error-prone environments.
ISPRS, 2024.
PDF ISPRS

Experience

Service

  • Reviewer for CVPR, ECCV, ICCV, ICLR, AVSS, ACCV, SIBGRAPI.

Awards

  • Robert E. Altenhofen Memorial Scholarship, American Society for Photogrammetry and Remote Sensing

Contact