Accurate and complete extraction of buildings from very high-resolution (VHR) remote sensing (RS) images is highly important for urban planning and land management. However, owing to the limited information available for small buildings and building boundaries, as well as challenges such as the spectral similarity of ground objects, tree occlusions, and shadow interference, automatically extracting buildings from VHR images remains challenging. These issues may result in building extraction errors such as misclassification, small building omissions, blurred boundaries, and incorrect segmentation. To address these challenges, we propose a multiscale hybrid transformer (MSHFormer) with boundary enhancement. This approach incorporates a hybrid encoder that combines a multiscale local perception (MSLP) module and a global perception module (GPM), combining the strengths of convolutional neural networks (CNNs) and transformers to achieve efficient synergy between global modeling and local feature extraction. In addition, we developed an edge enhancement module (EHM) to enhance boundary information, significantly improving building boundary segmentation accuracy. Finally, we design a group alignment feature fusion module (GAFFM) to efficiently integrate low-level features from the encoder with high-level features from the decoder, reducing feature space misalignment. The experimental results on three public datasets demonstrate the effectiveness of MSHFormer. Specifically, the proposed method achieves intersection-over-union (IoU) values of 89.1%, 73.6%, and 89.5% on the Potsdam, Massachusetts, and WHU datasets, respectively.

Nighttime light (NTL) data have been commonly used as a proxy for characterizing socioeconomic activities. Vegetation coverage has been found to be closely and inversely correlated with NTL intensity (NTLI). Although the combination of NTLI and vegetation indices has been studied in various applications, the complex relationship between NTLI and vegetation indices is not yet clear. By analyzing the relationship between NDVI and NTLI for the mainland China from 2013 to 2021, we found that the scatterplot between NDVI and NTLI exhibits a triangular shape with a physical meaning, which we called the NTLI-NDVI triangular feature space. Using the triangular feature space, we proposed the Vegetation Nighttime Condition Index (VNCI), which is defined as the ratio of NTLI differences among pixels with a specific NDVI value. VNCI is closely associated with local urban characteristics and has the ability to increase variation in NTLI conditions within urban areas. To demonstrate the application potential of the NTLI-NDVI triangular feature space and the proposed VNCI, two applications (urban area extraction and socioeconomic parameters estimation) were conducted. In terms of extracting urban areas, VNCI shows a better detection ability (with an average overall accuracy of 85.01%) than the original NPP-VIIRS NTL data and other two existing indices in extracting urban areas. Moreover, we further proposed a simple and novel NTL correcting approach to correct NTL using VNCI, which effectively eliminates the impact of vegetation and enhances the accuracy of estimating socioeconomic parameters. Our findings demonstrated that the VNCI-corrected NTL data show superior performance (with an average R2of 0.9) in estimating both the gross domestic product and electric power consumption at provincial level. We believe the NTLI-NDVI triangular feature space and VNCI hold great potential for NTL-based urban studies.

Deep neural networks suffer from performance decay when there is domain shift between the labeled source domain and unlabeled target domain, which motivates the research on domain adaptation (DA). Conventional DA methods usually assume that the labeled data is sampled from a single source distribution. However, in practice, labeled data may be collected from multiple sources, while naive application of the single-source DA algorithms may lead to suboptimal solutions. In this paper, we propose a novel multi-source distilling domain adaptation (MDDA) network, which not only considers the different distances among multiple sources and the target, but also investigates the different similarities of the source samples to the target ones. Specifically, the proposed MDDA includes four stages: (1) pre-train the source classifiers separately using the training data from each source; (2) adversarially map the target into the feature space of each source respectively by minimizing the empirical Wasserstein distance between source and target; (3) select the source training samples that are closer to the target to fine-tune the source classifiers; and (4) classify each encoded target feature by corresponding source classifier, and aggregate different predictions using respective domain weight, which corresponds to the discrepancy between each source and target. Extensive experiments are conducted on public DA benchmarks, and the results demonstrate that the proposed MDDA significantly outperforms the state-of-the-art approaches. Our source code is released at: https://github.com/daoyuan98/MDDA.

Input scale plays an important role in modern detection frameworks, and an optimal training scale for images exists empirically. However, the optimal one usually cannot be reached in facing extremely large images under the memory constraint. In this study, we explore the scale effect inside the object detection pipeline and find that feature upsampling with the introduction of high-resolution information benefits the detection. Compared with direct input upscaling, feature upsampling trades a small performance loss for a large amount of memory savings. From these observations, we propose a self-supervised feature augmentation network, which takes downsampled images as inputs and aims to generate comparable features with the ones when feeding upscaled images to networks. We present a guided feature upsampling module, which takes downsampled images as inputs, to learn upscaled feature representations with the supervision of real large features acquired from upscaled images. In a self-supervised learning manner, we can introduce detailed information of images to the network. For an efficient feature upsampling, we design a residualized sub-pixel convolution block based on a sub-pixel convolution layer, which involves considerable information in upsampling process. Experiments on Mapillary Vistas Dataset (MVD), Cityscapes, and COCO are conducted to demonstrate the effectiveness of our method. On the MVD and Cityscapes detection benchmarks, in which the images are extremely large, our method surpasses current approaches. On COCO, the proposed method obtains comparable results to existing methods but with higher efficiency.

Multi-shot person Re-IDentification (Re-ID) has recently received more research attention as its problem setting is more realistic compared to single-shot Re-ID in terms of application. While many large-scale single-shot Re-ID human image datasets have been released, most existing multishot Re-ID video sequence datasets containonly a few (i.e., several hundreds) human instances, which hinders further improvement of multi-shot Re-ID performance. To this end, we propose a deep cross-modality alignment network, which jointly explores both human sequence pairs and image pairs to facilitate training better multi-shot human Re-ID models, i.e., via transferring knowledge from image data to sequence data. To mitigate modality-to-modality mismatch issue, the proposed network is equipped with an image-to-sequence adaption module called cross-modality alignment sub-network, which successfully maps each human image into a pseudo human sequence to facilitate knowledge transferring and joint training. Extensive experimental results on several multi-shot person Re-ID benchmarks demonstrate great performance gain brought up by the proposed network.