Computer Vision

Boosting Vision Transformer robustness against attacks & noisy data, AdaNCA uses Neural Cellular Automata as plug-and-play adaptors between ViT layers, achieving significant accuracy improvement with …

Boosting Multi-Scene Pose Regression: Novel methods activate transformer self-attention, significantly improving camera pose estimation accuracy and efficiency.

TwinAct: Decoupling actions and actors for customizable text-guided action image generation.

ActFusion: a unified diffusion model achieving state-of-the-art performance in both action segmentation and anticipation by jointly learning visible and invisible parts of video sequences.

ActAnywhere, a novel video diffusion model, seamlessly integrates foreground subjects into new backgrounds by generating realistic video backgrounds tailored to subject motion, significantly reducing …

ACFun: A novel facial stylization method fusing abstract & concrete features for high-quality, artistically pleasing results from only one style & one face image.

This paper presents QSI-NMS and BOE-NMS, novel graph theory-based algorithms that significantly speed up non-maximum suppression (NMS) in object detection without significant accuracy loss, and introd…

Boost Vision Transformer pretraining speed by 4x with novel sequence compression techniques!

UniSeg3D: One model to rule them all! This unified framework masters six 3D segmentation tasks (panoptic, semantic, instance, interactive, referring, and open-vocabulary) simultaneously, outperforming…

Simple rule-based base-class mining (BCM) significantly boosts generalized few-shot semantic segmentation (GFSS) performance, surpassing complex existing methods.

UniDC framework achieves universal depth completion across various sensors and scenes using minimal labeled data, leveraging a foundation model and hyperbolic embedding for enhanced generalization.

Siamese Transformer with Hierarchical Refinement achieves state-of-the-art lane detection accuracy by integrating global and local features and a novel Curve-IoU loss.

This paper introduces a novel, robust inlier identification algorithm for point cloud registration that leverages lo-minimization.

GeCo: A novel single-stage low-shot counter achieving ~25% improvement in count accuracy, via unified object detection, segmentation, and counting.

A novel motion-aware spatio-temporal graph model surpasses existing methods in video salient object ranking by jointly optimizing multi-scale spatial and temporal features, thus accurately prioritizin…

A novel modular diffusion framework for image reconstruction dramatically cuts computational costs and achieves state-of-the-art perceptual quality across various tasks by cleverly combining pre-train…

Soft labels, not sophisticated data synthesis, are the key to successful dataset distillation, significantly improving data-efficient learning and challenging existing methods.

Depth-range-free MVS network using pose embedding achieves robust and accurate 3D reconstruction.

Researchers developed a lightweight protocol to probe large vision models’ 3D physical understanding by training classifiers on model features for various scene properties (geometry, material, lightin…

CAST: a novel consistency-aware spot-guided Transformer achieves state-of-the-art accuracy and efficiency in point cloud registration.