Joint Training Is Not Enough: Conditioned Cross-Granularity Training for Multimodal Document Understanding
研究通过条件跨粒度训练方法解决多模态文档理解中的细粒度和粗粒度任务互惠问题,相比联合训练在部分数据集上取得更好效果。
研究通过条件跨粒度训练方法解决多模态文档理解中的细粒度和粗粒度任务互惠问题,相比联合训练在部分数据集上取得更好效果。
研究通过12项任务和多种模型家族,揭示了输出格式对数据质量和模型能力评估的影响,并提出使用梯度特征等方法解决这一混淆问题。
为解决浅层决策树表达能力不足的问题,提出了一种多分支神经决策树(MBNDT),通过自适应剪枝和可微分多路分裂方法提高了预测准确性。
This work addresses the challenge of object detection in real-world scenarios where sparse annotations and unknown object categories coexist, leading to ambiguous supervision that hinders differentiation between unlabeled known objects and truly novel ones. To tackle this issue, the paper introduces Sparse Annotation Open-World Object Detection (SA-OWOD), a new task formulation, and proposes DPOD, a unified framework comprising a Known Target Recovery Module (KTRM) that reconstructs complete supervision signals and regularizes the feature space, and a Dual-view Discrepant Target Generator (DDTG) that leverages cross-view semantic discrepancies to identify reliable unknown objects. Evaluated on a sparse-annotation open-world benchmark, DPOD significantly outperforms existing methods, achieving notable gains especially in unknown object detection performance.
This work addresses the challenges of error accumulation and violation of degradation irreversibility in existing recursive health indicator prediction methods for long-term remaining useful life (RUL) estimation. To mitigate these issues, the authors propose a hybrid representation mechanism that integrates local and global features to reduce sensitivity to high-frequency noise. Additionally, they introduce a trend-guided recursive consistency loss function, which leverages soft dynamic time warping alignment and a multi-step rollout strategy to effectively bridge the gap between training and inference dynamics. Experimental results on two public bearing datasets demonstrate that the proposed approach significantly enhances the stability of long-term health indicator extrapolation and improves RUL prediction accuracy, while maintaining low computational complexity and strong generalization capability.
研究通过条件跨粒度训练方法解决多模态文档理解中的细粒度和粗粒度任务互惠问题,相比联合训练在部分数据集上取得更好效果。
研究通过12项任务和多种模型家族,揭示了输出格式对数据质量和模型能力评估的影响,并提出使用梯度特征等方法解决这一混淆问题。
为解决浅层决策树表达能力不足的问题,提出了一种多分支神经决策树(MBNDT),通过自适应剪枝和可微分多路分裂方法提高了预测准确性。
This work addresses the challenge of object detection in real-world scenarios where sparse annotations and unknown object categories coexist, leading to ambiguous supervision that hinders differentiation between unlabeled known objects and truly novel ones. To tackle this issue, the paper introduces Sparse Annotation Open-World Object Detection (SA-OWOD), a new task formulation, and proposes DPOD, a unified framework comprising a Known Target Recovery Module (KTRM) that reconstructs complete supervision signals and regularizes the feature space, and a Dual-view Discrepant Target Generator (DDTG) that leverages cross-view semantic discrepancies to identify reliable unknown objects. Evaluated on a sparse-annotation open-world benchmark, DPOD significantly outperforms existing methods, achieving notable gains especially in unknown object detection performance.
This work addresses the challenges of error accumulation and violation of degradation irreversibility in existing recursive health indicator prediction methods for long-term remaining useful life (RUL) estimation. To mitigate these issues, the authors propose a hybrid representation mechanism that integrates local and global features to reduce sensitivity to high-frequency noise. Additionally, they introduce a trend-guided recursive consistency loss function, which leverages soft dynamic time warping alignment and a multi-step rollout strategy to effectively bridge the gap between training and inference dynamics. Experimental results on two public bearing datasets demonstrate that the proposed approach significantly enhances the stability of long-term health indicator extrapolation and improves RUL prediction accuracy, while maintaining low computational complexity and strong generalization capability.