Sanqing Qu (瞿三清)

Hi there, welcome to my homepage! I am currently a postdoctoral researcher at Tongji University, where I am jointly supervised by Prof. Guang Chen and Prof. Changjun Jiang. I earned my Ph.D. from Tongji University in 2024.

My research focuses on transfer learning and its applications in Embodied AI and AI4Science.

Email  /  Google Scholar  /  Github

Selected Publications

* indicates equal contribution

We introduce RCP-Bench, the first comprehensive benchmark designed to evaluate the robustness of collaborative detection models under a wide range of real-world corruptions. RCPBench includes three new datasets (i.e., OPV2V-C, V2XSet-C, and DAIR-V2X-C) that simulate six collaborative cases and 14 types of camera corruption resulting from external environmental factors, sensor failures, and temporal misalignments. Extensive experiments on 10 leading collaborative perception models reveal that, while these models perform well under ideal conditions, they are significantly affected by corruptions.

We promote the vanilla BMD to BMD-v2 by incorporating a consistency-guided reweighting strategy to improve inter-class balanced sampling, and leveraging the silhouettes metric to realize adaptive intra-class multicentric clustering. Extensive experiments conducted on both 2D images and 3D point cloud recognition demonstrate that our proposed BMD strategy significantly improves existing representative methods.

Despite the simple global and local clustering (GLC) technique achieving commendable performance in separating "known" and "unknown" data, its reliance on pseudo-labeling supervision, especially using uniform encoding for all "unknown" data limits its capacity to discriminate among different "unknown" categories. To alleviate this, we promote GLC to GLC++ by developing a new contrastive affinity learning strategy, sidestepping the need for a specialized source model structure.

In this paper, we delve into TTA in 3D point cloud segmentation and propose a novel Hierarchical Geometry Learning (HGL) framework. HGL comprises three complementary modules from local, global to temporal learning in a bottom-up manner. Technically, we first construct a local geometry learning module for pseudo-label generation. Next, we build prototypes from the global geometry perspective for pseudo-label fine-tuning. Furthermore, we introduce a temporal consistency regularization module to mitigate negative transfer.

For Universal Domain Adaptation (UniDA), in this paper, we propose a new idea of LEArning Decomposition (LEAD), which decouples features into source-known and -unknown components to identify target-private data. This solution leads to elegant views for identifying target-private unknown data without tedious tuning thresholds or relying on iterative clustering.

In this paper, we target a practical setting in IP protection, i.e., Source-free Model IP protection. To achieve this, we propose a novel MAsk Pruning (MAP) framework. MAP stems from an intuitive hypothesis, i.e., there are target-related parameters in a well-trained model, locating and pruning them is the key to IP protection.

In this paper, we explore the Source-free Universal Domain Adaptation (SF-UniDA). SF-UniDA is appealing in view that universal model adaptation can be resolved only on the basis of a standard pre-trained closed-set model, i.e., without source raw data and dedicated model architecture.

Existing single-DG techniques commonly devise various data-augmentation algorithms, and remould the multi-source domain generalization methodology to learn domain-generalized (semantic) features. In contrast, we target a versatile Modality-Agnostic Debiasing (MAD) framework for single-DG, that enables generalization for different modalities.

In this paper, we proposed a bio-inspired event-camera based falls temporal localization framework. Specifically, we propose a event density-based action proposal generation scheme, and introduce a temporal-spatial attention mechanism for action modeling.

In this paper, we propose an action-context modeling network termed ACM-Net, which integrates a three-branch attention module to measure the likelihood of each temporal point being action instance, context, or non-action background, simultaneously.

Website Template