Text-to-image person re-identification (TIReID) aims to retrieve the most relevant pedestrian images from an image gallery based on natural language descriptions. Recent studies have achieved significant performance improvements by leveraging Masked Language Modeling (MLM) to align fine-grained information through local matching. However, in the text feature extraction, randomly masking text tokens may disrupt the semantic relationships between these local tokens, leading to feature misalignment; on the other hand, from an image feature perspective, redundant patches in pedestrian images hinder the information interaction across modalities. Moreover, the presence of noisy image-text pairs further complicates the learning process, as the model may be misled into recognizing incorrect patterns. To address these issues, we propose a robust fine-grained local alignment framework based on Key Phrase Dynamic Mask (KPDM). First, we strengthen the semantic relationships between text tokens by implementing a "adjective + noun'' phrase-level masking strategy, and design a frequency-based masked language loss (FMLM) to supervise fine-grained semantic-level local alignment. Second, we integrate cross-layer importance estimation to highlight key pedestrian image representations while removing redundant image features. Third, we propose a trusted consensus partitioning mechanism, utilizing intra-identity image-text similarity distributions to identify noisy pairs, enhancing the model robustness. Extensive experiments show that our method achieves 67.95% Rank-1 and 51.88% mAP on the RSTPReid dataset, exceeding the previous state-of-the-art by 2.6% and 1%. Furthermore, KPDM achieves Rank-1 accuracies of 75.97% on the CUHK-PEDES dataset and 67.78% on the ICFG-PEDES dataset, outperforming earlier methods.

Unsupervised domain adaptation for object detection aims to bridge the domain gap by transferring knowledge from a labeled source domain to an unlabeled target domain, thus improving the performance of detection models. Common strategies focus on aligning the feature distributions between source and target domains to reduce their discrepancies. However, achieving complete alignment is often not feasible in real-world situations due to a lack of annotations in the target domain. Recently, TeacherStudent approaches achieve feature alignment by generating reliable target pseudo-labels and become the dominant solution for addressing this issue. However, due to the domain shift, the teacher model bias to source domain, making it challenging to enhance the quality of target pseudo-labels. Some methods within this framework attempt to overcome the domain shift by incorporating distribution alignment components, yet these approaches also face challenges in achieving perfect alignment between domains. In this paper, we propose the Dual-Domain Teacher (DDT) method to address the domain adaptation detection problem by simultaneously detecting objects in both domains, thereby decreasing the need for perfect alignment. To address the issue of duplicate detection results produced by the Dual-Domain detection process, a candidate set refinement strategy is proposed to eliminate these duplicates across domains. Moreover, when teachers generate pseudo-labels by selecting reliable predictions with fixed confidence thresholds, valuable predictions may be overlooked in mutual learning. In our approach, a minimum variance-based dynamic threshold module is designed to mine valuable pseudo-labels by adaptively adjusting to the optimal threshold. Extensive experiments show that the DDT achieve a 56.7$\%$mAP on the CityScapes-to-Foggy CityScapes task, marking a 4.8 point improvement over the latest methods. On the PASCAL VOC-to-Clipart1k task, our method reaches 51.2$\%$mAP, outperforming previous state-of-the-art.

Anchor-free detection methods identify different objects by perceiving bounding box keypoints without predefined anchor boxes, which have attracted much attention due to their straightforward design and comparable performance. Currently, most anchor-free methods detect bounding box corners to regress object locations. In clutter environments, the bounding box corners may lie in background regions, which have limited relation with the object itself. In addition, the relationships between object keypoints are always neglected, potentially affecting the perceptibility of the detector for high-precision object detection. In this paper, we propose the Keypoint Relational Regression Network (KRRNet) to detect object keypoints with semantic relations instead of bounding box corners. The relational regression head is designed to enhance the keypoint relationship exploration capability and reason accuracy object locations. Moreover, the random background sampling strategy is proposed to sample negative background points around foreground object regions and form point pairs with object keypoints. Then, KRRNet can explicitly learn discriminative feature embedding from contrastive learning to pull close the positive pairs and push apart the negative pairs, resisting the influence of surrounding complex environments. KRRNet can be trained on one Nvidia RTX 3090 GPU and achieves a single-scale test AP of 48.9% and multi-scale test AP of 50.6% on the MS-COCO test-dev with the backbone of Hourglass-104, surpassing state-of-the-art bottom-up anchor-free detector using the same backbone.

We propose a novel two-stage cross-domain self-training (CDST) framework for unsupervised domain adaptive object detection in remote sensing. The first stage introduces the generative adversarial network (GAN)-based domain transfer strategy to preliminarily mitigate the domain shift for higher quality initial pseudo-labeled images, which utilizes the CycleGAN to transfer source-domain images to match the target domain. Moreover, the key issue in tailoring the self-training (ST) to unsupervised domain adaptive detection lies in the quality of pseudo-labeled images. To select high-quality pseudo-labeled images under the domain-shift circumstance, we propose hard example selection-based self-training (HES-ST) with the three key steps: 1) detector-based example division (DED), which divides the detected examples into easy examples and hard ones according to their confidence level; 2) confidence and relation joint score (CRJS)-based hard example selection, which combines two reliability levels calculated, respectively, by the detector and relation network (RN) module to mine reliable examples; and 3) union example (UE)-based training image selection, which combines both easy and reliable hard examples to choose target-domain images that may contain fewer detection errors. The experimental results on several remote sensing datasets demonstrate the effectiveness of our proposed framework. Compared with the baseline detector trained on the source dataset, our approach consistently improves the detection performance on the target dataset by 15.7%–16.8% mean average precision (mAP) and achieves the state-of-the-art (SOTA) results under various domain adaptation scenarios.

Automatic Modulation Classification (AMC) is a crucial task in the field of wireless communication, allowing for the identification of the modulation scheme of a received radio signal without prior knowledge of the communication system. Recently, AMC approaches based on Deep Learning (DL) have achieved outstanding results. However, the majority of current DL-based AMC methods face challenges in achieving high recognition accuracy while remaining computationally efficient. Some researchers have designed autoencoder-based models to generate low-dimensional temporal feature embeddings of the radio signal, thereby reducing the number of model parameters while maintaining high performance in recognizing modulation formats. However, when further improving AMC performance via learning low-dimensional spatial-temporal feature representations, traditional autoencoder models require both a convolutional decoder and an LSTM decoder to reconstruct temporal and spatial features separately, which unavoidably raises the model parameters. In this paper, we propose a spatiotemporal feature sharing reconstructing network (STARNet) to simultaneously extract low-dimensional spatial and temporal feature representations of radio signals using a single autoencoder structure, thereby reducing the number of model parameters and improving AMC performance. Additionally, we construct a Hybrid Attentive Ghost (HA-Ghost) to automatically extract discriminative radio signal spatial information according to signal reconstruction performance. Extensive experiments on benchmark datasets demonstrate that the proposed STARNet achieves an average modulation classification accuracy of 63.64%, outperforming previous state-of-the-art models. Despite extracting more types of features, STARNet has only 14,860 parameters, which is smaller than existing spatiotemporal autoencoder-based methods.