TLDR; We propose SignaVox, a sign-centered framework that scales 3D sign conversation data from isolated signs and enables direct sign-to-sign response generation.

TLDR; We propose a QK-tree framework that decomposes ViT attention into semantically meaningful query-key pair clusters, enabling interpretable analysis and training-free object localization across diverse ViT backbones.

TLDR; We propose QWERTY, a training-free motion control framework for image-to-video DiTs that warps semantic queries and uses self-guided latent optimization to achieve precise user-defined motion without fine-tuning.

TLDR; We propose ViKey, a training-free framework that enhances temporal reasoning in VideoLLMs by combining visual prompting with a Keyword-Frame Mapping (KFM) module, achieving dense-frame performance with only 20% of the frames.

TLDR; We introduce SlumpGuard, an AI-powered, real-time video analysis system for fully automated concrete slump prediction at construction sites, supported by a large-scale dataset of over 6,000 real-world videos of concrete discharge.

TLDR; We present MAVL, a multimodal benchmark for singable lyrics translation, and SylAVL-CoT, a model using audio-video cues and syllable constraints for natural, accurate results.

TLDR; We introduce VENUS, a large-scale video dataset for generating and understanding nonverbal expressions, along with MARS, a model designed to leverage it.

TLDR; We introduce ScalpVision, an AI system for comprehensive scalp disease and alopecia diagnosis that uses innovative hair segmentation and DiffuseIT-M, a generative model for dataset augmentation, to improve severity assessment and prediction accuracy.

TLDR; We introduce the effective preprocessing pipeline for sign language translation without glosses, combining skeleton-based motion features, keypoint normalization, and stochastic frame selection to enhance model performance.

TLDR; We introduce a two-stage vehicle class and orientation detection model using synthetic-to-real image translation and meta-table fusion to improve real-world prediction accuracy.

TLDR; (Korean) We introduce a real-time tram collision prediction system that combines fast object detection with YOLOv5 and a modified local dense optical flow to estimate object speed and predict collision time and probability using a single camera image.

TLDR; (Korean) We introduce a real-time pedestrian collision prediction system that uses YOLOv5 for fast object detection and a Local Dense Optical Flow method to quickly estimate pedestrian direction and speed, enabling accurate prediction of collision time and location.

TLDR; (Korean) We survey recent advances in optical flow estimation, comparing traditional and deep learning-based methods, and highlight their applications in autonomous driving, medical imaging, and surveillance systems.