Qiwei Yan
Current Organizations / Institutes
The organizations/institutes from the scholar's latest article
Research Subjects
Qiwei Yan has a total of 20 co-authors. Here are the top co-authors by number of shared publications:
- Chenglong Liu
Aerospace Information Research Institute, Chinese Academy of Sciences, Beijing, China
2 shared publications - Chubo Deng
Aerospace Information Research Institute, Chinese Academy of Sciences, Beijing, China
2 shared publications - Hongfeng Yu
Aerospace Information Research Institute, Chinese Academy of Sciences, Beijing, China
2 shared publications - Xian Sun
Chinese Academy of Sciences
2 shared publications - Zhongyan Hou
Key Laboratory of Target Cognition and Application Technology (TCAT) and the Aerospace Information Research Institute, Chinese Academy of Sciences, Beijing, China
2 shared publications - Chongyang Li
University of the Chinese Academy of Sciences, Beijing, China
1 shared publication
Latest Publications
- Open Access
- 10 November 2026
- Published by Association for Computing Machinery (ACM)
- p. 55-63
One Source, One Vote: Quotienting Multiplicity in Multimodal Retrieval
Multimodal retrieval represents an entity from a multiset of images and text, yet standard fusion gives repeated observations additional mass. Copies and re-encodings can thus alter representations and rankings and increase computation without adding an independent source. We call this failure representation replication bias. For a fixed source relation, we show that deterministic fusion is replication invariant precisely when it factors through the represented source classes. This yields an interface-aware rule: apply the quotient at the last stage where sources remain separable. With per-source embeddings, Replication-Invariant Fusion (RIF) weights the exact quotient through a Gaussian redundancy kernel; RIF-HG adds robust centers, equal group mass, and fixed modality mass for detected near copies. When a joint encoder hides per-view embeddings, Geometrically Verified Quotienting (GVQ) instead groups raw images before encoding. Controlled interventions on MCMR and FashionMV show that clean Recall can conceal rank instability. At matched clean MCMR Recall@10, RIF-HG lowers near-copy top-10 set replacement from 5.94% to 2.88% for SigLIP2 and from 3.53% to 1.26% for CLIP relative to nongrouped RIF-M. Its single-vector index is fivefold smaller and serves 3.77–3.99 × more queries per second than per-view max scoring. On FashionMV/ProCIR, GVQ uses 20.4% fewer visual tokens and halves genuine-pair merges relative to pHash, although uncached preprocessing is slower. A genuine-view counterfactual confirms that distinct views remain valuable, motivating joint evaluation of quality, stability, merge safety, and cost.
- 10 November 2026
- Published by Association for Computing Machinery (ACM)
- p. 55-63
Open AccessOne Source, One Vote: Quotienting Multiplicity in Multimodal Retrieval
Multimodal retrieval represents an entity from a multiset of images and text, yet standard fusion gives repeated observations additional mass. Copies and re-encodings can thus alter representations and rankings and increase computation without adding an independent source. We call this failure representation replication bias. For a fixed source relation, we show that deterministic fusion is replication invariant precisely when it factors through the represented source classes. This yields an interface-aware rule: apply the quotient at the last stage where sources remain separable. With per-source embeddings, Replication-Invariant Fusion (RIF) weights the exact quotient through a Gaussian redundancy kernel; RIF-HG adds robust centers, equal group mass, and fixed modality mass for detected near copies. When a joint encoder hides per-view embeddings, Geometrically Verified Quotienting (GVQ) instead groups raw images before encoding. Controlled interventions on MCMR and FashionMV show that clean Recall can conceal rank instability. At matched clean MCMR Recall@10, RIF-HG lowers near-copy top-10 set replacement from 5.94% to 2.88% for SigLIP2 and from 3.53% to 1.26% for CLIP relative to nongrouped RIF-M. Its single-vector index is fivefold smaller and serves 3.77–3.99 × more queries per second than per-view max scoring. On FashionMV/ProCIR, GVQ uses 20.4% fewer visual tokens and halves genuine-pair merges relative to pHash, although uncached preprocessing is slower. A genuine-view counterfactual confirms that distinct views remain valuable, motivating joint evaluation of quality, stability, merge safety, and cost.
- Open Access
- 9 November 2026
- Published by Association for Computing Machinery (ACM)
- p. 12969-12978
RAVA: Retrieval-Augmented Viewpoint Alignment for Subject-Driven Image Generation
- 9 November 2026
- Published by Association for Computing Machinery (ACM)
- p. 12969-12978
Open AccessRAVA: Retrieval-Augmented Viewpoint Alignment for Subject-Driven Image Generation
- 17 July 2025
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Transactions on Geoscience and Remote Sensing
- Vol. 63 (01962892) , 1-22
ReCon1M: A Large-Scale Benchmark Dataset for Relation Comprehension in Remote Sensing Imagery
Scene graph generation (SGG) is a high-level visual understanding and reasoning task aimed at extracting entities (such as objects) and their interrelationships from images. Significant progress has been made in the study of SGG in natural images in recent years, but its exploration in the domain of remote sensing images remains very limited. The complex characteristics of remote sensing images necessitate higher time and manual interpretation costs for annotation compared to natural images. The lack of a large-scale public SGG benchmark is a major impediment to the advancement of SGG-related research in aerial imagery. In this article, we introduce the first publicly available large-scale, million-level relation dataset in the field of remote sensing images, which is named ReCon1M. Specifically, our dataset is built upon FAIR1M and comprises 22 262 images. It includes annotations for 873 761 object bounding boxes across 60 categories and 1 052 223 relation triplets across 59 categories based on these bounding boxes. We provide a detailed description of the dataset’s characteristics and statistical information. In addition, an efficient global context-aware network (EGCAN) is proposed to improve inference efficiency in dense relation prediction through an object-pair pre-screening mechanism. By integrating visual, spatial, and semantic features, EGCAN captures fine-grained pairwise features and object-level contextual information to enhance its ability to discriminate relation. We conduct two object detection tasks and three subtasks within SGG on this dataset, assessing the performance of mainstream methods on these tasks. The experimental results show that the proposed EGCAN achieves state-of-the-art (SOTA) performance in 17 out of 24 accuracy metrics across three tasks and delivers the best performance in frames per second (FPS) for model inference. The ReCon1M dataset and related resources are available at https://recon1m-dataset.github.io/
- 17 July 2025
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Transactions on Geoscience and Remote Sensing
- Vol. 63 (01962892) , 1-22
ReCon1M: A Large-Scale Benchmark Dataset for Relation Comprehension in Remote Sensing Imagery
Scene graph generation (SGG) is a high-level visual understanding and reasoning task aimed at extracting entities (such as objects) and their interrelationships from images. Significant progress has been made in the study of SGG in natural images in recent years, but its exploration in the domain of remote sensing images remains very limited. The complex characteristics of remote sensing images necessitate higher time and manual interpretation costs for annotation compared to natural images. The lack of a large-scale public SGG benchmark is a major impediment to the advancement of SGG-related research in aerial imagery. In this article, we introduce the first publicly available large-scale, million-level relation dataset in the field of remote sensing images, which is named ReCon1M. Specifically, our dataset is built upon FAIR1M and comprises 22 262 images. It includes annotations for 873 761 object bounding boxes across 60 categories and 1 052 223 relation triplets across 59 categories based on these bounding boxes. We provide a detailed description of the dataset’s characteristics and statistical information. In addition, an efficient global context-aware network (EGCAN) is proposed to improve inference efficiency in dense relation prediction through an object-pair pre-screening mechanism. By integrating visual, spatial, and semantic features, EGCAN captures fine-grained pairwise features and object-level contextual information to enhance its ability to discriminate relation. We conduct two object detection tasks and three subtasks within SGG on this dataset, assessing the performance of mainstream methods on these tasks. The experimental results show that the proposed EGCAN achieves state-of-the-art (SOTA) performance in 17 out of 24 accuracy metrics across three tasks and delivers the best performance in frames per second (FPS) for model inference. The ReCon1M dataset and related resources are available at https://recon1m-dataset.github.io/
- 27 January 2025
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Transactions on Geoscience and Remote Sensing
- Vol. 63 (01962892) , 1-19
Hypergraph-Guided Multimodal Prototype for Remote Sensing Scene Understanding
Noticeable achievements have been made in entity-level perception tasks (e.g., object detection) in remote sensing image interpretation. But for remote sensing images carrying rich content, individual perception cannot well obtain the interaction patterns between entities. The recognition of relationships between entities is the key to deeply understanding remote sensing scenes. In this paper, we propose a hypergraph-guided multimodal prototype network (HMPNet), which performs relation recognition by matching relation representations with multimodal predicate prototypes. To overcome the imbalance of modal information in the matching process, a multimodal calibration strategy is devised, taking into account the image sub-prototype and text sub-prototype, which makes prediction results more reliable. Meanwhile, to align image and text sub-prototypes and explore relevant semantic patterns, the multimodal hypergraph is constructed to efficiently capture the associations between heterogeneous prototypes. Experimental results show that the performance of our model can reach the state-of-the-art (SOTA) level on the remote sensing scene graph generation task.
- 27 January 2025
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Transactions on Geoscience and Remote Sensing
- Vol. 63 (01962892) , 1-19
Hypergraph-Guided Multimodal Prototype for Remote Sensing Scene Understanding
Noticeable achievements have been made in entity-level perception tasks (e.g., object detection) in remote sensing image interpretation. But for remote sensing images carrying rich content, individual perception cannot well obtain the interaction patterns between entities. The recognition of relationships between entities is the key to deeply understanding remote sensing scenes. In this paper, we propose a hypergraph-guided multimodal prototype network (HMPNet), which performs relation recognition by matching relation representations with multimodal predicate prototypes. To overcome the imbalance of modal information in the matching process, a multimodal calibration strategy is devised, taking into account the image sub-prototype and text sub-prototype, which makes prediction results more reliable. Meanwhile, to align image and text sub-prototypes and explore relevant semantic patterns, the multimodal hypergraph is constructed to efficiently capture the associations between heterogeneous prototypes. Experimental results show that the performance of our model can reach the state-of-the-art (SOTA) level on the remote sensing scene graph generation task.