SkillNet: Create, Evaluate, and Connect AI Skills
为解决AI技能缺乏系统积累和转移的问题,提出SkillNet,一个创建、评估和组织AI技能的开放基础设施。
为解决AI技能缺乏系统积累和转移的问题,提出SkillNet,一个创建、评估和组织AI技能的开放基础设施。
Existing 3D skeleton-based action recognition research suffers from fragmented representation taxonomies and evaluation protocols misaligned with real-world scenarios; moreover, mainstream datasets lack critical dimensions—including rear-view perspectives, multi-person interactions, fine-grained or violent actions, and pandemic-era behaviors. To address these gaps, we propose a four-dimensional taxonomy (dataset design, spatial modeling, temporal modeling, and signal enhancement) and introduce ANUBIS: the first large-scale, multi-view 3D skeleton dataset explicitly designed for realistic challenges. ANUBIS features rear-view captures, 101 action classes (including 21 pandemic-related behaviors), and standardized recordings from 128 participants using Azure Kinect’s multi-sensor fusion. We further establish a unified benchmark framework, enabling reproducible evaluation of 12 state-of-the-art models. Our analysis identifies temporal modeling capacity and signal robustness as the primary bottlenecks limiting current performance.
Current single-image reflection removal (SIRR) research is hindered by the lack of large-scale, high-quality, real-world benchmark datasets. To address this, we introduce the first large-scale, in-the-wild SIRR benchmark—comprising 5,300 pixel-accurately aligned reflection/no-reflection image pairs—spanning diverse illumination conditions, object materials, and reflection patterns; it further includes 100 real-world, ground-truth-free images for generalization evaluation. A rigorously controlled acquisition pipeline ensures data fidelity. We propose an end-to-end U-Net-based removal model and comprehensively evaluate performance using five metrics: PSNR, SSIM, LPIPS, DISTS, and NIQE. Experiments confirm the dataset’s validity and establish a robust baseline. All data and code are publicly released to advance standardization and practical deployment of SIRR.
Real-world image degradations are diverse and challenging to identify automatically. Method: This paper pioneers the integration of vision-language models (VLMs) into degradation classification, proposing a fine-grained degradation-aware and modular restoration framework. It categorizes degradations into four types—super-resolution-related, reflection, motion blur, and no degradation—and employs a VLM to accurately classify each input image, subsequently activating a dedicated reconstruction network for on-demand enhancement. Domain adaptation is incorporated to improve cross-scenario generalization. Results: On multiple benchmark datasets, the method achieves 92.7% degradation classification accuracy and significantly outperforms unified enhancement approaches in PSNR and SSIM. It also enhances perceptual quality of restored images and improves performance on downstream tasks, thereby overcoming inherent limitations of end-to-end models in interpretability and generalizability.
Diffusion Transformers (DiTs) suffer from slow training convergence, and existing acceleration methods rely on external pretrained models, limiting their flexibility and generalization. This work proposes Self-Transcendence, a novel approach that dispenses with external semantic guidance and achieves fully self-supervised training by leveraging only internal model features. Specifically, it aligns shallow-layer DiT features with VAE latent representations and enhances semantic expressiveness of intermediate features through classifier-free guidance (CFG), enabling significant training acceleration using solely internal supervision signals. Experimental results demonstrate that, without any external pretrained models, the proposed method outperforms external-guidance approaches such as REPA in both training speed and generation quality.
为了解决语音识别中的隐式上下文问题,本文引入了HearInContext基准测试,通过同音词构建测试案例,并使用Qwen3-ASR-1.7B模型微调提高了目标召回率。
本文针对开放词汇多模态情绪识别中的模态缺失问题,提出了一种基于情感原型引导的融合框架APCF,通过构建情感原型库和条件检索来实现有效特征融合。
为解决长期记忆组织与演化问题,提出REALM框架,通过检索反馈驱动记忆重组,提高LLM代理的长期记忆性能。
本文提出一种无需训练的框架,通过结合布局交互先验、指令相关性和特征新颖性来解决GUI代理在轨迹展开中截图累积导致的推理延迟和内存使用问题。
该研究针对移动设备上跨源多粒度用户行为的表示学习问题,提出了一种新的预训练模型CM-PTM,通过层次级联掩码预测任务来解决数据稀疏性问题。