Generative Recommenders (GRs), exemplified by the Hierarchical Sequential Transduction Unit (HSTU), have emerged as a powerful paradigm for modeling long user interaction sequences. However, we observe that their ''flat-sequence'' assumption overlooks the rich, intrinsic structure of user behavior. This leads to two key limitations: a failure to capture the temporal hierarchy of session-based engagement, and computational inefficiency, as dense attention introduces significant noise that obscures true preference signals within semantically sparse histories, which deteriorates the quality of the learned representations. To this end, we propose a novel framework named HPGR (Hierarchical and Preference-aware Generative Recommender), built upon a two-stage paradigm that injects these crucial structural priors into the model to handle the drawback. Specifically, HPGR comprises two synergistic stages. First, a structure-aware pre-training stage employs a session-based Masked Item Modeling (MIM) objective to learn a hierarchically-informed and semantically rich item representation space. Second, a preference-aware fine-tuning stage leverages these powerful representations to implement a Preference-Guided Sparse Attention mechanism, which dynamically constrains computation to only the most relevant historical items, enhancing both efficiency and signal-to-noise ratio. Empirical experiments on a large-scale proprietary industrial dataset from APPGallery and an online A/B test verify that HPGR achieves state-of-the-art performance over multiple strong baselines, including HSTU and MTGR.

Few-shot object detection (FSOD) has been proposed to solve the problem of insufficient data for training, and it has drawn the attention of the remote sensing community in recent years. A mainstream type of FSOD method is to generate class prototypes based on the limited samples to help the construction of classification decision boundaries. However, these constructed prototypes may be far away from the true class centroids in the few-shot scenario. Recently, the vision-language model (VLM) has shown its powerful ability to align the visual features and text features, which leads to strong zero-shot performance on various downstream computer vision tasks when given only texts. Therefore, in this work, we propose to build class prototypes from text descriptions instead of limited visual instances by leveraging a classical pretrained VLM named CLIP. Concretely, we generate prototypes by feeding the CLIP text encoder with class names and enforcing each positive proposal feature to be close to the corresponding prototype. To accelerate the alignment process, we utilize the CLIP visual encoder as another teacher to achieve visual knowledge distillation. Moreover, we adopt prompt tuning to adapt CLIP to the remote sensing scenario. Extensive experiments on two public FSOD datasets, i.e., DIOR and NWPU VHR-10.v2, demonstrate the effectiveness of our method, which yields competitive results with that of existing approaches.

Real-world data ubiquitously exhibit long-tailed distribution, which sparks the increasing interest in long-tailed object detection (LTOD). However, existing methods neglect that a lack of diverse data in tail classes will cause underrepresented tail class features, making their efforts for balancing foreground classes tend to over-fit tail classes and be less effective. In this paper, we propose a multi-class co-attention generation network to increase data diversity of tail classes by generating augmented samples. To alleviate imbalance, we develop a distribution-aware up-sampling strategy, performing differential up-sampling for different classes and design a bi-directional regulation loss to adjust both positive and negative gradients. Moreover, we construct a new dataset LVIS-X with more rare classes based on existing LTOD benchmark dataset LVIS. Experiments on LVIS and LVIS-X demonstrate the superiority of the proposed method.

Video anomaly detection is to automatically identify predefined anomalous contents (e.g. abnormal objects, behaviors and scenes) in videos. The performance of video anomaly detection can be effectively improved by making the model focus more on the anomalous objects in videos. However, such existing approaches usually rely on pre-trained models, which not only require additional auxiliary information but also face the challenge of anomaly diversity in the real world. In this paper, we propose a new video anomaly detection method based on spatio-temporal relationships among objects. Concretely, we use a fully convolutional encoder-decoder network with symmetric skip connections as the backbone network, which can effectively extract features from the object regions at different scales. In the encoding stage, an attention mechanism is used to enhance the model’s understanding of the spatio-temporal relationships among various types of objects in the video. In the decoding stage, a dynamic pattern generator is designed to memorize the inter-object spatio-temporal relationships, which thus enhances the reconstructions of normal samples while making the reconstructions of abnormal samples more difficult. We conduct extensive experiments on three widely used video anomaly detection datasets CUHK Avenue , ShanghaiTech Campus and UCSD Ped2 , and the experimental results show that our proposed method can significantly improve the performance, and achieves state-of-the-art overall performance (considering both effectiveness and efficiency). In particular, our method achieves a state-of-the-art AUC of 98.4% on the UCSD Ped2 dataset that consists of various anomalies in real-world scenarios.

The generic object detection (GOD) task has been successfully tackled by recent deep neural networks, trained by an avalanche of annotated training samples from some common classes. However, it is still non-trivial to generalize these object detectors to the novel long-tailed object classes, which have only few labeled training samples. To this end, the Few-Shot Object Detection (FSOD) has been topical recently, as it mimics the humans’ ability of learning to learn and intelligently transfers the learned generic object knowledge from the common heavy-tailed to the novel long-tailed object classes. Especially, the research in this emerging field has been flourishing in recent years with various benchmarks, backbones, and methodologies proposed. To review these FSOD works, there are several insightful FSOD survey articles [ 58 , 59 , 74 , 78 ] that systematically study and compare them as the groups of fine-tuning/transfer learning and meta-learning methods. In contrast, we review the existing FSOD algorithms from a new perspective under a new taxonomy based on their contributions, i.e., data-oriented, model-oriented, and algorithm-oriented. Thus, a comprehensive survey with performance comparison is conducted on recent achievements of FSOD. Furthermore, we also analyze the technical challenges, the merits and demerits of these methods, and envision the future directions of FSOD. Specifically, we give an overview of FSOD, including the problem definition, common datasets, and evaluation protocols. The taxonomy is then proposed that groups FSOD methods into three types. Following this taxonomy, we provide a systematic review of the advances in FSOD. Finally, further discussions on performance, challenges, and future directions are presented.

Cross-view geo-localization is to match scene images (e.g. ground-view images) with geo-tagged aerial images, which is crucial to a wide range of applications such as autonomous driving and street view navigation. Existing methods can neither address the perspective difference well nor effectively capture the scene information. In this work, we propose a Geometric and Scene Information Fusion (GSIF) model for more accurate cross-view geo-localization. GSIF first learns the geometric information of scene images and aerial images via log-polar transformation and spatial-attention aggregation to alleviate the perspective difference. Then, it mines the scene information of scene images via Sky View Factor (SVF) extraction. Finally, both geometric information and scene information are fused for image matching, and a balanced loss function is introduced to boost the matching accuracy. Experimental results on two real datasets show that our model can significantly outperforms the existing methods.

Generic object detection has achieved significant progress in recent years. However, oriented object detection in aerial images is still a challenging task due to arbitrary orientation, complex backgrounds and large scale variation. Currently, the majority of oriented object detectors are anchor-based, achieving promising performance yet suffering from complicated anchor designs and imbalance between the positive and negative anchor boxes. In this work, we propose a new anchor-free model called RBA-CenterNet for oriented object detection. Specifically, we first detect the center point of each object after extracting feature of the input image. Then, we regress the rotated box parameters in other prediction branches. Considering that using only the center point of the object may hurt the detection performance, we introduce a refinement module called rotated border alignment (RBA) to integrate the border feature into the center point feature. Our experiments show that the proposed model RBA-CenterNet can achieve comparable detection performance to state-of-the-art methods.