Claims Browser

KDG-SemNOMA significantly outperforms state-of-the-art methods in both pixel-level accuracy and perceptual fidelity on FFHQ-256

"Experiments on FFHQ-256 demonstrate that KDG-SemNOMA significantly outperforms state-of-the-art methods in both pixel-level accuracy and perceptual fidelity."

Vision-Language Models demonstrate exceptional visual reasoning capabilities, but their inference costs escalate rapidly with the proliferation of visual tokens

"Vision-Language Models (VLMs) demonstrate exceptional visual reasoning capabilities, yet their inference costs escalate rapidly with the proliferation of visual tokens."

Existing visual token pruning methods fail to optimize the substantial latency of the visual encoding phase and often cannot jointly preserve holistic visual contexts and fine-grained details under strict token budgets

"Existing visual token pruning methods exhibit two fundamental limitations. First, most approaches operate exclusively post-vision encoder, leaving the substantial latency of the visual encoding phase unoptimized. Second, under strict token budgets, these methods often fail to jointly preserve holistic visual contexts and fine-grained details, leading to performance degradation."

PACE framework enables Qwen2.5-VL-7B to retain 93.8% of original performance while using only 10% of visual tokens, achieving 3.1x speedup in time to first token

"By integrating PACE into Qwen2.5-VL-7B, the model retains 93.8% of its original performance while utilizing only 10% of the visual tokens, yielding a 3.1x speedup in time to first token (TTFT)."

A training-free inference framework can accelerate both the vision encoder and LLM via a unified Condense-and-Extract paradigm that addresses visual token pruning limitations

"we propose PACE (Pixel-Adaptive Condense and Extract), a training-free inference framework that accelerates both the vision encoder and the Large Language Model (LLM) via a unified Condense-and-Extract paradigm."

Open World Object Detection systems built on multimodal foundation models suffer from semantic ambiguity due to unidirectional text-to-vision matching

"Open World Object Detection (OWOD) built on multimodal foundation models often suffers from semantic ambiguity caused by unidirectional text-to-vision matching"

Rigid outlier penalties in OWOD systems may over-suppress unknown objects near known-class decision boundaries

"rigid outlier penalties may over-suppress unknown objects near known-class decision boundaries"

CODE framework with OWL-ViT L/14 backbone achieves 21.7 U-mAP and 40.8 K-mAP in Task 1 of Real-World Detection benchmark, surpassing previous state of the art by 2.6 and 2.3 points respectively

"with the OWL-ViT L/14 backbone, CODE achieves 21.7 U-mAP and 40.8 K-mAP in Task 1, surpassing the previous state of the art by 2.6 and 2.3 points, respectively"

Cross-Modal Joint Confidence Calibration using global visual prototypes can improve text-driven known-class predictions in object detection

"Cross-Modal Joint Confidence Calibration injects global visual prototypes to calibrate text-driven known-class predictions"

Dynamic margin-aware outlier suppression preserves ambiguous out-of-distribution instances better than rigid suppression methods

"Dynamic Outlier Suppression via Confidence Margin replaces rigid suppression with a margin-aware adjustment that preserves ambiguous out-of-distribution instances"

A self-supervised framework can learn joint visuospatial representations from RGB-D observations that integrates depth-derived geometric priors with visual backbones

"We introduce a self-supervised framework for learning joint visuospatial representations from RGB-D observations."

Modern vision foundation models are trained almost exclusively on RGB images despite many embodied systems having access to explicit depth sensing that provides geometric information monocular inputs cannot recover

"While modern vision foundation models are trained almost exclusively on RGB images, many embodied systems have access to explicit depth sensing, which provides geometric information that monocular inputs cannot recover."

The RGB-D representation method outperforms prior methods of comparable scale on multiple 3D geometry benchmarks

"it outperforms prior methods of comparable scale on multiple 3D geometry benchmarks"

The resulting representation shows promising improvements on 3D awareness while preserving semantic transfer and remains competitive on RGB-D semantic segmentation tasks

"The resulting representation shows promising improvements on 3D awareness while preserving semantic transfer: it outperforms prior methods of comparable scale on multiple 3D geometry benchmarks, and remains competitive when probed for standard RGB-D semantic segmentation tasks."

The field of Human Activity Recognition using IMUs still lacks a unified model that can generalize across diverse subjects, devices, and activities

"the field still lacks a unified model that can generalize across diverse subjects, devices, and activities"

HALO outperforms five state-of-the-art baselines on all 8 aggregate metrics when evaluated on 7 held-out datasets

"HALO outperforms five state-of-the-art baselines on all 8 aggregate metrics"

HALO achieves better performance than MOMENT foundation model despite using 10x fewer parameters (35M vs 341.2M)

"Despite using only ~35M trainable parameters -- 10x fewer than the latest foundation model MOMENT (341.2M) -- HALO improves zero-shot open-set accuracy"

HALO improves zero-shot open-set accuracy by 13.7 percentage points over MOMENT across 87 training labels

"HALO improves zero-shot open-set accuracy, measured over all 87 training labels, by 13.7 percentage points"