Selected work

Representative Papers

TeaMatch: Teachable Cross-Modal Representation Learning for 2D-3D Matching

Aug 10, 2026

This work addresses the challenge of establishing robust 2D–3D correspondences under severe degradation conditions such as noise, low overlap, and structural ambiguity. The authors propose TeaMatch, a novel framework that introduces teachability into cross-modal representation learning for the first time. By employing a task-oriented weak student to simulate typical failure modes and optimizing representations to recover the teacher’s features, TeaMatch enhances structural consistency and robustness. The approach integrates a teacher–student architecture with correspondence-level constraints and geometry-aware regularization, seamlessly fitting into coarse-to-fine matching pipelines without incurring additional inference overhead. Extensive experiments demonstrate that TeaMatch achieves state-of-the-art performance across multiple challenging 2D–3D matching benchmarks, significantly improving matching robustness.

0 citationsRead paper

SWINSleepNet: A Hierarchical Context-Aware Framework for Sleep Staging (v2)

Aug 03, 2026

This study addresses the limited performance of existing automatic sleep staging methods in ambiguous and transitional stages—particularly N1—due to insufficient modeling of fine-grained intra-epoch structures and cross-regional spectral dependencies. To overcome this, the authors propose a dual-stream hierarchical context-aware framework that jointly processes raw time-domain EEG signals and multi-scale time-frequency representations. Convolutional encoders capture waveform details, while Swin Transformers model local spectro-temporal features and long-range dependencies. A bidirectional context module then fuses multi-branch features to explicitly refine both intra-epoch representations and inter-epoch temporal relationships. Evaluated on the Sleep-EDF-20/78 and SHHS datasets, the method achieves state-of-the-art performance, significantly improving robustness and accuracy in identifying N1 and other transitional sleep stages.

0 citationsRead paper

SHReg: Strictly Rotation-Equivariant Point Cloud Registration via Spherical Harmonics

Jul 25, 2026

This work addresses the limited robustness and discriminability of local features under arbitrary 3D rotations in point cloud registration by proposing the first strictly rotation-equivariant registration framework that operates without a local reference frame. Built upon SO(3) representation theory, the method employs spherical harmonics to construct a rotation-equivariant neural network that jointly learns rotation-invariant descriptors and equivariant geometric features. This design enables each putative correspondence to directly model the underlying rigid transformation, substantially reducing reliance on extensive RANSAC sampling. Experiments on the 3DMatch, 3DLoMatch, and KITTI benchmarks demonstrate that the proposed approach achieves significantly higher registration accuracy than existing methods under large rotational perturbations.

0 citationsRead paper

Personalized News Recommendation with Multi-granularity Candidate-aware User Modeling

Apr 19, 2025

Existing news recommendation methods predominantly rely on static user representations derived solely from click behaviors, failing to capture users’ diverse interests and overlooking multi-granularity associations between candidate news and user preferences. To address these limitations, we propose a multi-granularity candidate-aware user modeling framework. Our approach introduces, for the first time, a candidate-driven three-level attention mechanism—operating at the word, entity, and news levels—to jointly model fine-grained, dynamic, and context-sensitive relevance between candidate news and user interests. The architecture integrates a news text encoder with a knowledge-enhanced entity extractor and incorporates a dedicated multi-granularity feature fusion network. Extensive experiments on real-world datasets demonstrate that our method achieves a 1.82% improvement in AUC over state-of-the-art baselines, validating the effectiveness of multi-granularity candidate awareness in enhancing recommendation accuracy.

0 citationsRead paper

Recent publications

Latest Papers

TeaMatch: Teachable Cross-Modal Representation Learning for 2D-3D Matching

Aug 10, 2026

This work addresses the challenge of establishing robust 2D–3D correspondences under severe degradation conditions such as noise, low overlap, and structural ambiguity. The authors propose TeaMatch, a novel framework that introduces teachability into cross-modal representation learning for the first time. By employing a task-oriented weak student to simulate typical failure modes and optimizing representations to recover the teacher’s features, TeaMatch enhances structural consistency and robustness. The approach integrates a teacher–student architecture with correspondence-level constraints and geometry-aware regularization, seamlessly fitting into coarse-to-fine matching pipelines without incurring additional inference overhead. Extensive experiments demonstrate that TeaMatch achieves state-of-the-art performance across multiple challenging 2D–3D matching benchmarks, significantly improving matching robustness.

0 citationsRead paper

SWINSleepNet: A Hierarchical Context-Aware Framework for Sleep Staging (v2)

Aug 03, 2026

This study addresses the limited performance of existing automatic sleep staging methods in ambiguous and transitional stages—particularly N1—due to insufficient modeling of fine-grained intra-epoch structures and cross-regional spectral dependencies. To overcome this, the authors propose a dual-stream hierarchical context-aware framework that jointly processes raw time-domain EEG signals and multi-scale time-frequency representations. Convolutional encoders capture waveform details, while Swin Transformers model local spectro-temporal features and long-range dependencies. A bidirectional context module then fuses multi-branch features to explicitly refine both intra-epoch representations and inter-epoch temporal relationships. Evaluated on the Sleep-EDF-20/78 and SHHS datasets, the method achieves state-of-the-art performance, significantly improving robustness and accuracy in identifying N1 and other transitional sleep stages.

0 citationsRead paper

SHReg: Strictly Rotation-Equivariant Point Cloud Registration via Spherical Harmonics

Jul 25, 2026

This work addresses the limited robustness and discriminability of local features under arbitrary 3D rotations in point cloud registration by proposing the first strictly rotation-equivariant registration framework that operates without a local reference frame. Built upon SO(3) representation theory, the method employs spherical harmonics to construct a rotation-equivariant neural network that jointly learns rotation-invariant descriptors and equivariant geometric features. This design enables each putative correspondence to directly model the underlying rigid transformation, substantially reducing reliance on extensive RANSAC sampling. Experiments on the 3DMatch, 3DLoMatch, and KITTI benchmarks demonstrate that the proposed approach achieves significantly higher registration accuracy than existing methods under large rotational perturbations.

0 citationsRead paper

Personalized News Recommendation with Multi-granularity Candidate-aware User Modeling

Apr 19, 2025

Existing news recommendation methods predominantly rely on static user representations derived solely from click behaviors, failing to capture users’ diverse interests and overlooking multi-granularity associations between candidate news and user preferences. To address these limitations, we propose a multi-granularity candidate-aware user modeling framework. Our approach introduces, for the first time, a candidate-driven three-level attention mechanism—operating at the word, entity, and news levels—to jointly model fine-grained, dynamic, and context-sensitive relevance between candidate news and user interests. The architecture integrates a news text encoder with a knowledge-enhanced entity extractor and incorporates a dedicated multi-granularity feature fusion network. Extensive experiments on real-world datasets demonstrate that our method achieves a 1.82% improvement in AUC over state-of-the-art baselines, validating the effectiveness of multi-granularity candidate awareness in enhancing recommendation accuracy.

0 citationsRead paper