Selected work

Representative Papers

Egoistic MDS-based Rigid Body Localization

Jan 20, 2025

This paper addresses the problem of prior-free relative pose estimation between unknown-shape rigid bodies in autonomous driving. We propose an anchor-free rigid-body localization method that relies solely on one-sided ranging measurements. Our key contribution is the first integration of multidimensional scaling (MDS) with a bi-centering operator into rigid-body kinematic modeling—eliminating the conventional requirement of geometric shape consistency between the two bodies and enabling ego-centric pose estimation for arbitrarily shaped, heterogeneous rigid bodies. The method constructs an MDS framework from the distance matrix and jointly optimizes it under rigid-body kinematic constraints. Simulation results demonstrate significant reductions in RMSE for both translational and rotational pose estimates across diverse rigid-body configurations. The approach exhibits strong robustness to measurement noise and generalizes effectively across disparate geometric morphologies.

1 citationsRead paper

Streamlining the Development of Active Learning Methods in Real-World Object Detection

Aug 27, 2025

In real-world object detection, active learning development faces two major bottlenecks: prohibitively high computational cost (up to 282 GPU-hours per detector training) and unreliable evaluation (large performance ranking fluctuations across validation sets). To address these, we propose Object-level Set Similarity (OSS), the first object-level feature-based similarity metric that unifies active learning training and evaluation without requiring detector training. OSS quantifies sample selection quality and identifies ineffective strategies directly from object features. It further enables construction of representative validation sets, significantly improving evaluation stability. Evaluated on KITTI, BDD100K, and CODA, OSS is compatible with mainstream detectors—including EfficientDet and YOLOv3—and accurately predicts method performance trends. Each OSS-based evaluation saves up to 282 GPU-hours, drastically reducing development cost while enhancing reliability—particularly critical for safety-critical applications.

0 citationsRead paper

Following the Clues: Experiments on Person Re-ID using Cross-Modal Intelligence

Jul 02, 2025

Street-scene person re-identification (Re-ID) faces dual challenges of privacy leakage—particularly from non-facial personally identifiable information (PII) in open datasets—and limited cross-domain robustness. Method: We propose cRID, the first systematic framework leveraging semantic cues to detect non-facial PII in images. It integrates vision-language models to extract text-describable sensitive semantic features, employs a graph attention network to model fine-grained PII correlations, and applies interpretable representation learning for privacy-aware feature disentanglement. Contribution/Results: Evaluated on cross-dataset benchmarks (e.g., Market-1501 → CUHK03-np [detected]), cRID achieves significant mAP improvements over baselines, demonstrating its effectiveness in enhancing Re-ID generalization while rigorously preserving privacy. The framework offers both theoretical insight into semantic PII modeling and practical utility for privacy-compliant surveillance systems.

0 citationsRead paper

Index Modulated Affine Frequency Division Multiplexing With Spread Spectrum

May 14, 2025

To address the severe performance degradation of AFDM systems over doubly-dispersive (time–frequency jointly spread) channels under high-mobility conditions, this paper proposes IM-AFDM-SS—a novel energy-efficient waveform integrating index modulation (IM) and direct-sequence spread spectrum (DSSS). Its key innovation lies in the first joint modulation of spreading code indices and AFDM subcarrier indices, enabling simultaneous enhancement of spectral and energy efficiency. A low-complexity two-stage maximum-ratio combining (MRC) detection algorithm is designed, and a tight bit-error-rate (BER) upper bound under maximum-likelihood (ML) detection is analytically derived. Theoretical analysis and simulations demonstrate that, compared to AFDM-SS and IM-AFDM, the proposed IM-AFDM-SS achieves significantly lower BER in high-mobility scenarios, while offering superior robustness against doubly-dispersive interference and a more favorable energy–efficiency trade-off.

0 citationsRead paper

Closing the Loop: Motion Prediction Models beyond Open-Loop Benchmarks

May 08, 2025

Motion prediction models often exhibit improved open-loop accuracy without corresponding gains in closed-loop driving performance, revealing a critical misalignment between standard evaluation metrics and real-world autonomy. Method: This paper introduces the first systematic prediction-planning co-evaluation framework for closed-loop assessment, integrating state-of-the-art learning-based predictors (e.g., Transformer- and GAN-based models) with realistic planners within end-to-end CARLA and nuScenes-based driving simulations. Contribution/Results: We empirically demonstrate that open-loop prediction accuracy is not a reliable proxy for closed-loop performance; instead, temporal consistency of predictions and compatibility with downstream planners are decisive factors. Notably, a lightweight model with 86% fewer parameters achieves superior closed-loop driving metrics. The work establishes a new paradigm for prediction-planning co-evaluation and releases open-source code to advance closed-loop–oriented motion prediction research.

0 citationsRead paper

Recent publications

Latest Papers

Streamlining the Development of Active Learning Methods in Real-World Object Detection

Aug 27, 2025

In real-world object detection, active learning development faces two major bottlenecks: prohibitively high computational cost (up to 282 GPU-hours per detector training) and unreliable evaluation (large performance ranking fluctuations across validation sets). To address these, we propose Object-level Set Similarity (OSS), the first object-level feature-based similarity metric that unifies active learning training and evaluation without requiring detector training. OSS quantifies sample selection quality and identifies ineffective strategies directly from object features. It further enables construction of representative validation sets, significantly improving evaluation stability. Evaluated on KITTI, BDD100K, and CODA, OSS is compatible with mainstream detectors—including EfficientDet and YOLOv3—and accurately predicts method performance trends. Each OSS-based evaluation saves up to 282 GPU-hours, drastically reducing development cost while enhancing reliability—particularly critical for safety-critical applications.

0 citationsRead paper

Following the Clues: Experiments on Person Re-ID using Cross-Modal Intelligence

Jul 02, 2025

Street-scene person re-identification (Re-ID) faces dual challenges of privacy leakage—particularly from non-facial personally identifiable information (PII) in open datasets—and limited cross-domain robustness. Method: We propose cRID, the first systematic framework leveraging semantic cues to detect non-facial PII in images. It integrates vision-language models to extract text-describable sensitive semantic features, employs a graph attention network to model fine-grained PII correlations, and applies interpretable representation learning for privacy-aware feature disentanglement. Contribution/Results: Evaluated on cross-dataset benchmarks (e.g., Market-1501 → CUHK03-np [detected]), cRID achieves significant mAP improvements over baselines, demonstrating its effectiveness in enhancing Re-ID generalization while rigorously preserving privacy. The framework offers both theoretical insight into semantic PII modeling and practical utility for privacy-compliant surveillance systems.

0 citationsRead paper

Index Modulated Affine Frequency Division Multiplexing With Spread Spectrum

May 14, 2025

To address the severe performance degradation of AFDM systems over doubly-dispersive (time–frequency jointly spread) channels under high-mobility conditions, this paper proposes IM-AFDM-SS—a novel energy-efficient waveform integrating index modulation (IM) and direct-sequence spread spectrum (DSSS). Its key innovation lies in the first joint modulation of spreading code indices and AFDM subcarrier indices, enabling simultaneous enhancement of spectral and energy efficiency. A low-complexity two-stage maximum-ratio combining (MRC) detection algorithm is designed, and a tight bit-error-rate (BER) upper bound under maximum-likelihood (ML) detection is analytically derived. Theoretical analysis and simulations demonstrate that, compared to AFDM-SS and IM-AFDM, the proposed IM-AFDM-SS achieves significantly lower BER in high-mobility scenarios, while offering superior robustness against doubly-dispersive interference and a more favorable energy–efficiency trade-off.

0 citationsRead paper

Closing the Loop: Motion Prediction Models beyond Open-Loop Benchmarks

May 08, 2025

Motion prediction models often exhibit improved open-loop accuracy without corresponding gains in closed-loop driving performance, revealing a critical misalignment between standard evaluation metrics and real-world autonomy. Method: This paper introduces the first systematic prediction-planning co-evaluation framework for closed-loop assessment, integrating state-of-the-art learning-based predictors (e.g., Transformer- and GAN-based models) with realistic planners within end-to-end CARLA and nuScenes-based driving simulations. Contribution/Results: We empirically demonstrate that open-loop prediction accuracy is not a reliable proxy for closed-loop performance; instead, temporal consistency of predictions and compatibility with downstream planners are decisive factors. Notably, a lightweight model with 86% fewer parameters achieves superior closed-loop driving metrics. The work establishes a new paradigm for prediction-planning co-evaluation and releases open-source code to advance closed-loop–oriented motion prediction research.

0 citationsRead paper

On Background Bias of Post-Hoc Concept Embeddings in Computer Vision DNNs

Apr 11, 2025

This work identifies and quantifies an implicit background dependency in posterior concept-embedding C-XAI methods for visual DNNs: such methods frequently misattribute background statistical shortcuts as semantic concepts, leading to explanation failure under atypical backgrounds (e.g., animals on roads). To address this, the authors propose the first systematic diagnostic framework—an scalable evaluation paradigm based on background randomization—integrating a Net2Vec variant for concept activation analysis and a multi-concept benchmark spanning 50+ concepts, two datasets, and seven model architectures. Empirical results provide the first large-scale evidence of severe background bias in mainstream C-XAI methods. A key finding is that lightweight background perturbations substantially improve both concept robustness and segmentation generalizability. This discovery establishes a practical, theoretically grounded pathway toward developing background-robust concept-based explanations.

0 citationsRead paper