Paper-Conference

Exaone-4.0-VL, a 32B multi-modal foundation model, achieves SOTA across vision-language benchmarks and empowers enterprise AI agent applications.

SO(3)-equivariant pose harmonics enable consistent single-image 3D orientation estimation by directly regressing Wigner-D coefficients in the frequency domain, overcoming discontinuities in spatial parameterizations.

This work proposes a self-supervised framework using group-equivariant CNNs and a novel group-aligning technique to learn discriminative, rotation-invariant local descriptors for robust image matching and pose estimation.

This work introduces a self-supervised framework with rotation-equivariant CNNs and a dense orientation alignment loss to detect robust oriented keypoints, achieving state-of-the-art results in image matching and camera pose estimation.

This work presents a self-supervised histogram alignment framework for estimating scale and orientation of image patches, achieving superior pose estimation and boosting image matching and 6-DoF camera pose estimation.

This work proposes a feature distillation framework that learns compact yet effective local descriptors by combining multi-level CNN features, achieving up to 100× smaller size while maintaining state-of-the-art performance across image matching tasks.

This work proposes Hyperpixel Flow, a real-time matching algorithm using hyperpixels that combine multi-level CNN features with Hough voting, achieving state-of-the-art correspondence on three benchmarks and the large-scale SPair-71k dataset.

This work introduces Dynamic Hyperpixel Flow, a method that adaptively composes hypercolumn features from relevant CNN layers conditioned on image pairs, achieving state-of-the-art semantic correspondence with improved efficiency.

This work introduces an attentive semantic alignment method with an offset-aware correlation kernel that filters out distractors and captures local transformations, achieving state-of-the-art performance in semantic correspondence.