Selected work

Representative Papers

Good Deep Features to Track: Self-Supervised Feature Extraction and Tracking in Visual Odometry

Sep 10, 2025

To address unstable feature detection and tracking in large-scale, long-term outdoor visual odometry (VO) caused by illumination variations, dynamic scenes, and low-texture regions, this paper proposes a task-driven self-supervised feature learning framework. Unlike existing supervised approaches relying on SuperPoint/SuperGlue, our method employs VO motion estimation error as a feedback signal to iteratively optimize feature detection, description, and matching in an end-to-end self-supervised training paradigm. This closed-loop optimization significantly improves feature robustness and out-of-distribution generalization. Experimental results demonstrate that the proposed method enhances feature tracking stability by 23.6% and improves VO localization accuracy by 18.4% on challenging real-world sequences—particularly excelling in low-texture and strongly varying illumination conditions.

0 citationsRead paper

Local Search Improvements for Soft Happy Colouring

Jun 23, 2025

This study addresses the soft ρ-happy coloring problem on partially colored graphs: given a fixed precoloring of vertices, find a vertex coloring that maximizes the number of ρ-happy vertices—those whose fraction of same-color neighbors is at least ρ. The problem is NP-hard and closely related to community detection. We propose three efficient local search algorithms, including one with provable linear-time complexity. We provide the first systematic analysis revealing a positive correlation between ρ and community detection accuracy: higher ρ values yield ρ-happy colorings that more faithfully reflect ground-truth communities, albeit at the cost of increased difficulty in achieving balanced color partitions. Extensive experiments demonstrate that our algorithms significantly outperform state-of-the-art methods in both ρ-happy vertex coverage and computational efficiency. This work establishes a new paradigm for unsupervised community detection via graph coloring and delivers practical, scalable tools for this task.

0 citationsRead paper

Privacy-Aware Lifelong Learning

May 16, 2025International Conference on Learning Representations

This work addresses privacy-compliant continual learning, tackling the core challenge of *how to accurately forget knowledge of specific tasks—thereby satisfying the “right to be forgotten”—while continually acquiring new tasks, avoiding catastrophic forgetting, enabling forward knowledge transfer, and minimizing memory overhead*. We propose the first unified framework integrating continual learning with machine unlearning, featuring a task-level precise unloading mechanism: it employs task-specific sparse subnetworks for parameter isolation and sharing, augmented by lightweight episodic memory replay to jointly ensure privacy preservation, learning stability, and computational efficiency. Evaluated on multiple image classification benchmarks, our method significantly outperforms existing approaches, achieving state-of-the-art privacy-aware continual learning performance. It is the first to enable verifiable, controllable, and low-overhead task-level knowledge addition and deletion within a single neural network.

0 citationsRead paper

Adversarially Robust Spiking Neural Networks with Sparse Connectivity

May 16, 2025

To address the challenge of simultaneously achieving adversarial robustness, memory efficiency, and energy efficiency for deep neural networks on resource-constrained embedded systems, this paper proposes a novel sparse spiking neural network (SNN) construction method. Our approach uniquely enables *co-transfer* of both adversarial robustness and structural sparsity from robust artificial neural networks (ANNs): we first obtain a robust and sparse ANN via adversarial training, then design a new conversion algorithm to map its sparse connectivity and weights to an SNN, incorporating sparse connection modeling and lightweight adversarial fine-tuning. Experiments demonstrate that, compared to dense SNNs, our method maintains high clean accuracy and significantly improves adversarial robustness—e.g., +23.5% robust accuracy against PGD attacks on CIFAR-10—while reducing weight storage by 100× and improving energy efficiency by 8.6×.

0 citationsRead paper

ISP-AD: A Large-Scale Real-World Dataset for Advancing Industrial Anomaly Detection with Synthetic and Real Defects

Mar 06, 2025

Industrial anomaly detection is hindered by the scarcity of high-quality datasets that faithfully represent real-world defect appearances and complex imaging conditions. To address this, we introduce ISP-AD—the first large-scale, real-world defect dataset tailored for screen printing scenarios—featuring minute, low-contrast defects synthesized under high-design-variability backgrounds and collected directly from production lines. We propose a scalable training paradigm: “synthetic pretraining + incremental injection of real samples,” integrating GAN- and rendering-based synthetic defects with a small number of authentic defects to enable hybrid-supervised optimization. Experiments demonstrate that incorporating only a minimal set of real defects substantially enhances model generalization—achieving high recall while maintaining low false-positive rates. This effectively overcomes the generalization bottleneck of existing methods under challenging imaging conditions, advancing anomaly detection toward zero-defect manufacturing.

0 citationsRead paper

Recent publications

Latest Papers

Good Deep Features to Track: Self-Supervised Feature Extraction and Tracking in Visual Odometry

Sep 10, 2025

To address unstable feature detection and tracking in large-scale, long-term outdoor visual odometry (VO) caused by illumination variations, dynamic scenes, and low-texture regions, this paper proposes a task-driven self-supervised feature learning framework. Unlike existing supervised approaches relying on SuperPoint/SuperGlue, our method employs VO motion estimation error as a feedback signal to iteratively optimize feature detection, description, and matching in an end-to-end self-supervised training paradigm. This closed-loop optimization significantly improves feature robustness and out-of-distribution generalization. Experimental results demonstrate that the proposed method enhances feature tracking stability by 23.6% and improves VO localization accuracy by 18.4% on challenging real-world sequences—particularly excelling in low-texture and strongly varying illumination conditions.

0 citationsRead paper

Local Search Improvements for Soft Happy Colouring

Jun 23, 2025

This study addresses the soft ρ-happy coloring problem on partially colored graphs: given a fixed precoloring of vertices, find a vertex coloring that maximizes the number of ρ-happy vertices—those whose fraction of same-color neighbors is at least ρ. The problem is NP-hard and closely related to community detection. We propose three efficient local search algorithms, including one with provable linear-time complexity. We provide the first systematic analysis revealing a positive correlation between ρ and community detection accuracy: higher ρ values yield ρ-happy colorings that more faithfully reflect ground-truth communities, albeit at the cost of increased difficulty in achieving balanced color partitions. Extensive experiments demonstrate that our algorithms significantly outperform state-of-the-art methods in both ρ-happy vertex coverage and computational efficiency. This work establishes a new paradigm for unsupervised community detection via graph coloring and delivers practical, scalable tools for this task.

0 citationsRead paper

Privacy-Aware Lifelong Learning

May 16, 2025International Conference on Learning Representations

This work addresses privacy-compliant continual learning, tackling the core challenge of *how to accurately forget knowledge of specific tasks—thereby satisfying the “right to be forgotten”—while continually acquiring new tasks, avoiding catastrophic forgetting, enabling forward knowledge transfer, and minimizing memory overhead*. We propose the first unified framework integrating continual learning with machine unlearning, featuring a task-level precise unloading mechanism: it employs task-specific sparse subnetworks for parameter isolation and sharing, augmented by lightweight episodic memory replay to jointly ensure privacy preservation, learning stability, and computational efficiency. Evaluated on multiple image classification benchmarks, our method significantly outperforms existing approaches, achieving state-of-the-art privacy-aware continual learning performance. It is the first to enable verifiable, controllable, and low-overhead task-level knowledge addition and deletion within a single neural network.

0 citationsRead paper

Adversarially Robust Spiking Neural Networks with Sparse Connectivity

May 16, 2025

To address the challenge of simultaneously achieving adversarial robustness, memory efficiency, and energy efficiency for deep neural networks on resource-constrained embedded systems, this paper proposes a novel sparse spiking neural network (SNN) construction method. Our approach uniquely enables *co-transfer* of both adversarial robustness and structural sparsity from robust artificial neural networks (ANNs): we first obtain a robust and sparse ANN via adversarial training, then design a new conversion algorithm to map its sparse connectivity and weights to an SNN, incorporating sparse connection modeling and lightweight adversarial fine-tuning. Experiments demonstrate that, compared to dense SNNs, our method maintains high clean accuracy and significantly improves adversarial robustness—e.g., +23.5% robust accuracy against PGD attacks on CIFAR-10—while reducing weight storage by 100× and improving energy efficiency by 8.6×.

0 citationsRead paper

ISP-AD: A Large-Scale Real-World Dataset for Advancing Industrial Anomaly Detection with Synthetic and Real Defects

Mar 06, 2025

Industrial anomaly detection is hindered by the scarcity of high-quality datasets that faithfully represent real-world defect appearances and complex imaging conditions. To address this, we introduce ISP-AD—the first large-scale, real-world defect dataset tailored for screen printing scenarios—featuring minute, low-contrast defects synthesized under high-design-variability backgrounds and collected directly from production lines. We propose a scalable training paradigm: “synthetic pretraining + incremental injection of real samples,” integrating GAN- and rendering-based synthetic defects with a small number of authentic defects to enable hybrid-supervised optimization. Experiments demonstrate that incorporating only a minimal set of real defects substantially enhances model generalization—achieving high recall while maintaining low false-positive rates. This effectively overcomes the generalization bottleneck of existing methods under challenging imaging conditions, advancing anomaly detection toward zero-defect manufacturing.

0 citationsRead paper