Domain Gap Research Articles

Cross-View Image Synthesis From a Single Image With Progressive Parallel GAN

Cross-view image synthesis aims to synthesize a ground-view image covering the same geographic region for a given single aerial-view image (or vice versa). Existing approaches typically tackle this challenging task by relaxing the single-image constraint and using a ground-truth semantic map as additional input to aid synthesis. However, this is nearly infeasible in practice. In this paper, we investigate how to generate a detail-enriched and structurally accurate ground-level image from only a single aerial-level input image, in which there are no other prior knowledge except for the input image. Towards this goal, we propose a novel Progressive Parallel Generative Adversarial Network (PPGAN) that starts from generating low-resolution outputs and progressively produces ground images at higher resolutions as the network propagates forward. In this manner, our PPGAN decomposes the task into several manageable sub-tasks, which helps to generate detail-enriched and structurally accurate ground images. During progressive generation, the PPGAN employs a parallel generation paradigm that enables the generator to produce multi-resolution images in parallel, thereby avoiding excessive time cost on training. Furthermore, for effective information propagation across multi-resolution images, a feature fusion module (FFM) is devised to mitigate the domain gap between cross-level image features, which enables a balance of detail and structural information synthesis. Additionally, the proposed Channel-Space Attention Selection Module (CSASM) learns the mapping relationship between cross-view images in a larger scale space to enhance the quality of the output image. Quantitative and qualitative experiments demonstrate that, our method requires only one input image without the aid of additional inputs, but is capable of synthesizing detail-enriched and structurally accurate ground images and outperforms the existing state-of-the-art methods on two famous benchmarks.

Domain Adaptation for Underwater Image Enhancement.

Recently, learning-based algorithms have shown impressive performance in underwater image enhancement. Most of them resort to training on synthetic data and obtain outstanding performance. However, these deep methods ignore the significant domain gap between the synthetic and real data (i.e., inter-domain gap), and thus the models trained on synthetic data often fail to generalize well to real-world underwater scenarios. Moreover, the complex and changeable underwater environment also causes a great distribution gap among the real data itself (i.e., intra-domain gap). However, almost no research focuses on this problem and thus their techniques often produce visually unpleasing artifacts and color distortions on various real images. Motivated by these observations, we propose a novel Two-phase Underwater Domain Adaptation network (TUDA) to simultaneously minimize the inter-domain and intra-domain gap. Concretely, in the first phase, a new triple-alignment network is designed, including a translation part for enhancing realism of input images, followed by a task-oriented enhancement part. With performing image-level, feature-level and output-level adaptation in these two parts through jointly adversarial learning, the network can better build invariance across domains and thus bridging the inter-domain gap. In the second phase, an easy-hard classification of real data according to the assessed quality of enhanced images is performed, in which a new rank-based underwater quality assessment method is embedded. By leveraging implicit quality information learned from rankings, this method can more accurately assess the perceptual quality of enhanced images. Using pseudo labels from the easy part, an easy-hard adaptation technique is then conducted to effectively decrease the intra-domain gap between easy and hard samples. Extensive experimental results demonstrate that the proposed TUDA is significantly superior to existing works in terms of both visual quality and quantitative metrics.

Multi-Modal Cross-Domain Alignment Network for Video Moment Retrieval

As an increasingly popular task in multimedia information retrieval, video moment retrieval (VMR) aims to localize the target moment from an untrimmed video according to a given language query. Most previous methods depend heavily on numerous manual annotations ( <italic xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">i.e.</i> , moment boundaries), which are extremely expensive to acquire in practice. In addition, due to the domain gap between different datasets, directly applying these pre-trained models to an unseen domain leads to a significant performance drop. In this paper, we focus on a novel task: cross-domain VMR, where fully-annotated datasets are available in one domain (“source domain”), but the domain of interest (“target domain”) only contains unannotated datasets. As far as we know, we present the first study on cross-domain VMR. To address this new task, we propose a novel Multi-Modal Cross-Domain Alignment (MMCDA) network to transfer the annotation knowledge from the source domain to the target domain. However, due to the domain discrepancy between the source and target domains and the semantic gap between videos and queries, directly applying trained models to the target domain generally leads to a performance drop. To solve this problem, we develop three novel modules: (i) a domain alignment module is designed to align the feature distributions between different domains of each modality; (ii) a cross-modal alignment module aims to map both video and query features into a joint embedding space and to align the feature distributions between different modalities in the target domain; (iii) a specific alignment module tries to obtain the fine-grained similarity between a specific frame and the given query for optimal localization. By jointly training these three modules, our MMCDA can learn domain-invariant and semantic-aligned cross-modal representations. Extensive experiments on three challenging benchmarks (ActivityNet Captions, Charades-STA and TACoS) illustrate that our cross-domain method MMCDA outperforms all state-of-the-art single-domain methods. Impressively, MMCDA raises the performance by more than 7% in representative cases, which demonstrates its effectiveness.

Crowd Counting via Unsupervised Cross-Domain Feature Adaptation

Given an image, crowd counting aims to estimate the amount of target objects in the image. With un-predictable installation situations of surveillance systems (or other equipment), crowd counting images from different data sets may exhibit severe discrepancies in viewing angle, scale, lighting condition, etc. As it is usually expensive and time-consuming to annotate each data set for model training, it has been an essential issue in crowd counting to transfer a well-trained model on a labeled data set (source domain) to a new data set (target domain). To tackle this problem, we propose a cross-domain learning network to learn the domain gaps in an unsupervised learning manner. The proposed network comprises of a Multi-granularity Feature-aware Discriminator (MFD) module, a Domain-Invariant Feature Adaptation (DFA) module, and a Cross-domain Vanishing Bridge (CVB) module to remove domain-specific information from the extracted features and promote the mapping performances of the network. Unlike most existing methods that use only Global Feature Discriminator (GFD) to align features at image level, an additional Local Feature Discriminator (LFD) is inserted and together with GFD form the MFD module. As a complement to MFD, LFD refines features at pixel level and has the ability to align local features. The DFA module explicitly measures the distances between the source domain features and the target domain features and aligns the marginal distribution of their features with Maximum Mean Discrepancy (MMD). Finally, the CVB module provides an incremental capability of removing the impact of interfering part of the extracted features. Several well-known networks are adopted as the backbone of our algorithm to prove the effectiveness of the proposed adaptation structure. Comprehensive experiments demonstrate that our model achieves competitive performance to the state-of-the-art methods. Code and pre-trained models are available at https://github.com/gcding/CDFA-pytorch.

Source-Free Active Domain Adaptation via Augmentation-Based Sample Query and Progressive Model Adaptation.

Active domain adaptation (ADA), which enormously improves the performance of unsupervised domain adaptation (UDA) at the expense of annotating limited target data, has attracted a surge of interest. However, in real-world applications, the source data in conventional ADA are not always accessible due to data privacy and security issues. To alleviate this dilemma, we introduce a more practical and challenging setting, dubbed as source-free ADA (SFADA), where one can select a small quota of target samples for label query to assist the model learning, but labeled source data are unavailable. Therefore, how to query the most informative target samples and mitigate the domain gap without the aid of source data are two key challenges in SFADA. To address SFADA, we propose a unified method SQAdapt via augmentation-based Sample Query and progressive model Adaptation. In specific, an active selection module (ASM) is built for target label query, which exploits data augmentation to select the most informative target samples with high predictive sensitivity and uncertainty. Then, we further introduce a classifier adaptation module (CAM) to leverage both the labeled and unlabeled target data for progressively calibrating the classifier weights. Meanwhile, the source-like target samples with low selection scores are taken as source surrogates to realize the distribution alignment in the source-free scenario by the proposed distribution alignment module (DAM). Moreover, as a general active label query method, SQAdapt can be easily integrated into other source-free UDA (SFUDA) methods, and improve their performance. Comprehensive experiments on multiple benchmarks have shown that SQAdapt can achieve superior performance and even surpass most of the ADA methods.

Transformer-Based Synthetic-to-Measured SAR Image Translation via Learning of Representational Features

Deep-learning-based target recognition in synthetic aperture radar (SAR) images has been actively studied in recent years. However, it is very costly to collect large numbers of labeled SAR images, especially measured SAR target images of various classes, to train high-performance classification networks. To solve the problem of insufficient SAR data, electromagnetic computational tools have often been developed and used to synthesize the measured SAR target images from data modeling. However, despite the use of sophisticated SAR image modeling, there is a large domain gap between synthetic SAR images and measured images such that networks trained with synthetic SAR images tend to show poor classification performance when tested on measured SAR target images. In this paper, we propose a novel transformer-based synthetic-to-measured SAR target image translation network, referred to as SAR-SMT Net, to bridge the gap between synthetic and measured SAR target images. SAR-SMT Net takes synthetic SAR target images as input and estimates the latent representational features of their corresponding measured SAR images to faithfully adjust the global context and scattering characteristics of the input synthetic SAR target images to the corresponding measured SAR values. In addition, we propose five challenging experimental scenarios that can validate SAR image translation performance outcomes. Experimentally, SAR-SMT Net as proposed here outperforms previous state-of-the-art methods in the experiment scenarios, demonstrating feasible generalization ability when used to translate synthetic SAR target images into their corresponding measured SAR target images with a high level of fidelity, even for unseen target classes at unseen azimuth angles.

Memory-Contrastive Unsupervised Domain Adaptation for Building Extraction of High-Resolution Remote Sensing Imagery

Deep learning-based semantic segmentation has been widely applied for building extraction. However, due to the domain gap, the extraction of building in high-resolution remote sensing imagery is difficult when the model trained on a source dataset is directly used to test on a target data. Considering that humans can retrieve memory to deal with correlative tasks in different domains, memory mechanisms have been developed effectively to assist cross-domain feature extraction. However, whether the memory mechanisms can achieve satisfactory result or not highly depends on the premise that the memory is relevant to the task. Therefore, the domain-invariant memory is crucial in cross-domain building extraction task. To this end, a memory-contrastive unsupervised domain adaptation method is proposed on the basis of a novel memory mechanism. Specifically, to facilitate the model to memorize domain-invariant features, we first conduct a normalization-based image style transfer strategy and a discriminator-based adversarial method at the image level and feature level, respectively. Subsequently, we carry out a memory-contrastive module to obtain domain-invariant features. Especially, a teacher–student network is exploited to help knowledge transferring by knowledge distillation to enhance the performance of the memory-contracted module. To narrow the distance between the two domains, a memory bank is designed to store and update category features obtained from the source domain, and then the similarity between category features in the target domain and memory bank is calculated. Results of the cross-domain experiments show that the proposed method can achieve optimal building extraction. (GitHub: https://github.com/RS-CSU/MDANet).

Learning Adaptive Patch Generators for Mask-Robust Image Inpainting

In this paper, we propose a Mask-Robust Inpainting Network (MRIN) method to recover the masked areas of an image. Most existing methods learn a single model for image inpainting, under a basic assumption that all the masks belong to the same type. However, we discover that the masks are usually complex and exhibit various shapes and sizes at different locations of an image, where a single model cannot fully capture the large domain gap across different masks. To address this, we learn to decompose a complex mask area into several basic mask types and inpaint the damaged image in a patch-wise manner with a type-specific generator. More specifically, our MRIN consists of a mask-robust agent and an adaptive patch generative network. The mask-robust agent contains a mask selector and a patch locator, which generates mask attention maps to select a patch at each step. We train our mask-robust agent to learn the optimal inpainting patch route in a reinforcement learning manner by formulating the process of inpainting sequentially as a Markov decision process. Then, based on the predicted mask attention maps, the adaptive patch generative network inpaints the selected patch with the generators bank, so that it sequentially inpaints each patch with different patch generators according to its mask type. Extensive experiments demonstrate that our approach outperforms most state-of-the-art approaches on the Place2, CelebA, and Paris Street View datasets.

Learning Shape-Invariant Representation for Generalizable Semantic Segmentation.

Semantic segmentation assigns a category for each pixel and has achieved great success in a supervised manner. However, it fails to generalize well in new domains due to the domain gap. Domain adaptation is a popular way to solve this issue, but it needs target data and cannot handle unavailable domains. In domain generalization (DG), the model is trained without the target data and DG aims to generalize well in new unavailable domains. Recent works reveal that shape recognition is beneficial for generalization but still lack exploration in semantic segmentation. Meanwhile, the object shapes also exist a discrepancy in different domains, which is often ignored by the existing works. Thus, we propose a Shape-Invariant Learning (SIL) framework to focus on learning shape-invariant representation for better generalization. Specifically, we first define the structural edge, which considers both the object boundary and the inner structure of the object to provide more discrimination cues. Then, a shape perception learning strategy including a texture feature discrepancy reduction loss and a structural feature discrepancy enlargement loss is proposed to enhance the shape perception ability of the model by embedding the structural edge as a shape prior. Finally, we use shape deformation augmentation to generate samples with the same content and different shapes. Essentially, our SIL framework performs implicit shape distribution alignment at the domain-level to learn shape-invariant representation. Extensive experiments show that our SIL framework achieves state-of-the-art performance.

Face De-Occlusion With Deep Cascade Guidance Learning

Occlusion is a challenging yet commonly seen problem for facial perception. Existing works resort to deep learning models and perform model training on synthesized data due to the lack of paired real-world data. As a result, they usually perform unsatisfactorily on real-world occluded faces because of domain gaps. In this paper, we decompose the face de-occlusion task into three stages, i.e., occlusion detection, face parsing, and face reconstruction, to alleviate this issue. We first perform occlusion detection and use its results as guidance for the second stage to conduct occlusion-free face parsing. As such, face deocclusion is first performed on the face paring space with less difficulty. We can train these two stages on both synthesized and real-world images, hence can obtain accurate results for the latter. In the last stage, we use the domain-agnostic occlusion detection map and the face parsing map as the guidance to conduct face reconstruction, thus can reduce the impact of appearance information and improve the model performance on real-world data. Aiming at improving the model capacity of inferring occluded facial appearance, we also propose two types of reference modules to use relevant facial parts to enhance the reconstruction of occluded regions. Consequently, our proposed model achieves promising face de-occlusion results on real-world images.

SCAN++: Enhanced Semantic Conditioned Adaptation for Domain Adaptive Object Detection

Domain Adaptive Object Detection (DAOD) transfers an object detector from the labeled source domain to a novel unlabelled target domain. Recent advances bridge the domain gap by aligning category-agnostic feature distribution and minimizing the domain discrepancy for adapting semantic distribution. Though great success, these methods model domain discrepancy with prototypes within a batch, yielding a biased estimation of domain-level statistics. Moreover, the category-agnostic alignment leads to the disagreement of the cross-domain semantic distribution with inevitable classification errors. To address these two issues, we propose an enhanced Semantic Conditioned AdaptatioN (SCAN++) framework, which leverages unbiased semantics for DAOD. Specifically, in the source domain, we design the conditional kernel to sample Pixel of Interests (PoIs), and aggregate PoIs with a cross-image graph to estimate an unbiased semantic sequence. Conditioned on the semantic sequence, we further update the parameter of the conditional kernel in a semantic conditioned manifestation module, and establish a novel conditional graph in the target domain to model unlabeled semantics. After modeling the semantic distribution in both domains, we integrate the conditional kernel into adversarial alignment to achieve semantic-aware adaptation in a Conditional Kernel guided Alignment (CKA) module. Meanwhile, the Semantic Sequence guided Transport (SST) module is proposed to transfer reliable semantic knowledge to the target domain through solving the cross-domain Optimal Transport (OT) assignment, achieving unbiased adaptation at the semantic level. Comprehensive experiments on four adaptation scenarios demonstrate that SCAN++ achieves state-of-the-art results. The code is available at <uri xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">https://github.com/CityU-AIM-Group/SCAN/tree/SCAN++</uri> .

DAKRS: Domain Adaptive Knowledge-Based Retrieval System for Natural Language-Based Vehicle Retrieval

Given Natural Language (NL) text descriptions, NL-based vehicle retrieval aims to extract target vehicles from a multi-view multi-camera traffic video pool. Due to inherent distinctions between textual and visual data, this is a challenging multi-modal retrieval task that requires robust feature extractors (e.g. neural network) to well-align the abstract representations of texts and images in the same domain. However, solutions to the problem have been challenged by the high data complexities of not only the multi-view, multi-camera attributes of visual data and the diverse range of textual descriptions but also a lack of high-volume datasets in this relatively new field, alongside a prominently large domain gap between training and test sets. Many existing approaches have developed computationally expensive models to separately extract the subspaces of language and vision before blending into the same shared representation space while only focusing on single-modal information and ignoring much of the multi-modal information to deal with the aforementioned issues. Hence, we propose a Domain Adaptive Knowledge-based Retrieval System (DAKRS) to effectively and efficiently align multi-modal knowledge in a setting of limited labels. Our contributions are threefold: (i) An efficient extension of Contrastive Language-Image Pre-training (CLIP)’s transfer learning into a baseline text-to-image multi-modular vehicle retrieval framework; (ii) A data enhancement module to create pseudo-vehicle tracks from the traffic video pool by leveraging the robustness of baseline retrieval model combine with background subtraction; and (iii) A SSDA (SSDA) scheme to engineer pseudo-labels for adapting model parameters to the target domain distribution. Experimental results are benchmarked on the Cityflow-NL dataset, illustrating our competitiveness against state-of-the-art performances in terms of effectiveness and efficiency without needing further post-processing or ensembling.

An Empirical Study of Remote Sensing Pretraining

Deep learning has largely reshaped remote sensing (RS) research for aerial image understanding and made a great success. Nevertheless, most of the existing deep models are initialized with the ImageNet pretrained weights. Since natural images inevitably present a large domain gap relative to aerial images, probably limiting the finetuning performance on downstream aerial scene tasks. This issue motivates us to conduct an empirical study of remote sensing pretraining (RSP) on aerial images. To this end, we train different networks from scratch with the help of the largest RS scene recognition dataset up to now &#x2014; MillionAID, to obtain a series of RS pretrained backbones, including both convolutional neural networks (CNN) and vision transformers such as Swin and ViTAE, which have shown promising performance on computer vision tasks. Then, we investigate the impact of RSP on representative downstream tasks including scene recognition, semantic segmentation, object detection, and change detection using these CNN and vision transformer backbones. Empirical study shows that RSP can help deliver distinctive performances in scene recognition tasks and in perceiving RS related semantics such as &#x201C;Bridge&#x201D; and &#x201C;Airplane&#x201D;. We also find that, although RSP mitigates the data discrepancies of traditional ImageNet pretraining on RS images, it may still suffer from task discrepancies, where downstream tasks require different representations from scene recognition tasks. These findings call for further research efforts on both large-scale pretraining datasets and effective pretraining methods. The codes and pretrained models will be released at https://github.com/ViTAE-Transformer/ViTAE-Transformer-Remote-Sensing.

Pan-tumor T-lymphocyte detection using deep neural networks: Recommendations for transfer learning in immunohistochemistry

The success of immuno-oncology treatments promises long-term cancer remission for an increasing number of patients. The response to checkpoint inhibitor drugs has shown a correlation with the presence of immune cells in the tumor and tumor microenvironment. An in-depth understanding of the spatial localization of immune cells is therefore critical for understanding the tumor’s immune landscape and predicting drug response. Computer-aided systems are well suited for efficiently quantifying immune cells in their spatial context. Conventional image analysis approaches are often based on color features and therefore require a high level of manual interaction. More robust image analysis methods based on deep learning are expected to decrease this reliance on human interaction and improve the reproducibility of immune cell scoring. However, these methods require sufficient training data and previous work has reported low robustness of these algorithms when they are tested on out-of-distribution data from different pathology labs or samples from different organs. In this work, we used a new image analysis pipeline to explicitly evaluate the robustness of marker-labeled lymphocyte quantification algorithms depending on the number of training samples before and after being transferred to a new tumor indication. For these experiments, we adapted the RetinaNet architecture for the task of T-lymphocyte detection and employed transfer learning to bridge the domain gap between tumor indications and reduce the annotation costs for unseen domains. On our test set, we achieved human-level performance for almost all tumor indications with an average precision of 0.74 in-domain and 0.72–0.74 cross-domain. From our results, we derive recommendations for model development regarding annotation extent, training sample selection, and label extraction for the development of robust algorithms for immune cell scoring. By extending the task of marker-labeled lymphocyte quantification to a multi-class detection task, the pre-requisite for subsequent analyses, e.g., distinguishing lymphocytes in the tumor stroma from tumor-infiltrating lymphocytes, is met.

Depth-Assisted ResiDualGAN for Cross-Domain Aerial Images Semantic Segmentation

Unsupervised domain adaptation (UDA) is an approach to minimizing the domain gap. Generative methods are common approaches to minimizing the domain gap of aerial images, which improves the performance of the downstream tasks, for example, cross-domain semantic segmentation. For aerial images, the digital surface model (DSM) is usually available in both the source domain and the target domain. Depth information in DSM brings external information to generative models. However, little research utilizes it. In this letter, depth-assisted ResiDualGAN (DRDG) is proposed where depth supervised loss (DSL) and depth cycle consistency loss (DCCL) are used to bring depth information into the generative model. Experimental results show that DRDG reaches state-of-the-art accuracy between generative methods in cross-domain semantic segmentation tasks. Source code is available at <uri xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">https://github.com/miemieyanga/ResiDualGAN-DRDG</uri> .

Cross-Modality Pyramid Alignment for Visual Intention Understanding.

Visual intention understanding is the task of exploring the potential and underlying meaning expressed in images. Simply modeling the objects or backgrounds within the image content leads to unavoidable comprehension bias. To alleviate this problem, this paper proposes a Cross-modality Pyramid Alignment with Dynamic optimization (CPAD) to enhance the global understanding of visual intention with hierarchical modeling. The core idea is to exploit the hierarchical relationship between visual content and textual intention labels. For visual hierarchy, we formulate the visual intention understanding task as a hierarchical classification problem, capturing multiple granular features in different layers, which corresponds to hierarchical intention labels. For textual hierarchy, we directly extract the semantic representation from intention labels at different levels, which supplements the visual content modeling without extra manual annotations. Moreover, to further narrow the domain gap between different modalities, a cross-modality pyramid alignment module is designed to dynamically optimize the performance of visual intention understanding in a joint learning manner. Comprehensive experiments intuitively demonstrate the superiority of our proposed method, outperforming existing visual intention understanding methods.

Semi-Self-Supervised Learning for Semantic Segmentation in Images with Dense Patterns.

Deep learning has shown potential in domains with large-scale annotated datasets. However, manual annotation is expensive, time-consuming, and tedious. Pixel-level annotations are particularly costly for semantic segmentation in images with dense irregular patterns of object instances, such as in plant images. In this work, we propose a method for developing high-performing deep learning models for semantic segmentation of such images utilizing little manual annotation. As a use case, we focus on wheat head segmentation. We synthesize a computationally annotated dataset-using a few annotated images, a short unannotated video clip of a wheat field, and several video clips with no wheat-to train a customized U-Net model. Considering the distribution shift between the synthesized and real images, we apply three domain adaptation steps to gradually bridge the domain gap. Only using two annotated images, we achieved a Dice score of 0.89 on the internal test set. When further evaluated on a diverse external dataset collected from 18 different domains across five countries, this model achieved a Dice score of 0.73. To expose the model to images from different growth stages and environmental conditions, we incorporated two annotated images from each of the 18 domains to further fine-tune the model. This increased the Dice score to 0.91. The result highlights the utility of the proposed approach in the absence of large-annotated datasets. Although our use case is wheat head segmentation, the proposed approach can be extended to other segmentation tasks with similar characteristics of irregularly repeating patterns of object instances.

Cross-Domain Facial Expression Recognition via Contrastive Warm up and Complexity-Aware Self-Training.

Unsupervised cross-domain Facial Expression Recognition (FER) aims to transfer the knowledge from a labeled source domain to an unlabeled target domain. Existing methods strive to reduce the discrepancy between source and target domain, but cannot effectively explore the abundant semantic information of the target domain due to the absence of target labels. To this end, we propose a novel framework via Contrastive Warm up and Complexity-aware Self-Training (namely CWCST), which facilitates source knowledge transfer and target semantic learning jointly. Specifically, we formulate a contrastive warm up strategy via features, momentum features, and learnable category centers to concurrently learn discriminative representations and narrow the domain gap, which benefits domain adaptation by generating more accurate target pseudo labels. Moreover, to deal with the inevitable noise in pseudo labels, we develop complexity-aware self-training with a label selection module based on prediction entropy, which iteratively generates pseudo labels and adaptively chooses the reliable ones for training, ultimately yielding effective target semantics exploration. Furthermore, by jointly using the two mentioned components, our framework enables to effectively utilize the source knowledge and target semantic information by source-target co- training. In addition, our framework can be easily incorporated into other baselines with consistent performance improvements. Extensive experimental results on seven databases show the superior performance of the proposed method against various baselines.

Domain Adaptive Remote Sensing Scene Recognition via Semantic Relationship Knowledge Transfer

Scene recognition has attracted rising attentions of many researchers in the remote sensing fields, owing to the rapidly advancing of remote sensing devices in recent years. However, images obtained from various sensors dominate diverse sensor-specific characteristics, which will dramatically weaken the model transferability trained on a source data domain to a different target domain on account of the domain shift issues. To mitigate the domain discrepancy, most existing methods attend to align the cross-domain distributions. While the valuable knowledge of semantic relationships between different scenes is generally overlooked, and the underlying correlation across scenes cannot be fully discovered. For the sake of tackling this challenge, we propose an adaptive remote sensing scene recognition network, which can successfully transfer both the discriminative knowledge and cross-scene relationship from source to target. Specifically, in this paper, we acquire sensor-invariant representations in an adversarial manner and realize fine-grained conditional distribution alignment contrastively. In such a way, the tremendous domain gap can be mitigated to a large extent, and the discriminative and well-matched representations will be derived favorably. In addition, we explicitly construct class-wise relationship distributions belonging to two domains respectively and minimize their divergence to conduct semantic relationship knowledge transfer (SRKT), for the purpose of sufficiently unearthing the intrinsic semantic relative structures that can prompt generality of the model in the target domain. Finally, we conduct multiple experiments on representative multi-domain remote sensing benchmarks, and the extensive experimental results demonstrate the superiority of our proposed approach.

Learning Domain and Pose Invariance for Thermal-to-Visible Face Recognition

Interest in thermal to visible face recognition has grown significantly over the last decade due to advancements in thermal infrared cameras and analytics beyond the visible spectrum. Despite large discrepancies between thermal and visible spectra, existing approaches bridge domain gaps by either synthesizing visible faces from thermal faces or by learning the cross-spectrum image representations. These approaches typically work well with frontal facial imagery collected at varying ranges and expressions, but exhibit significantly reduced performance when matching thermal faces with varying poses to frontal visible faces. We propose a novel Domain and Pose Invariant Framework that simultaneously learns domain and pose invariant representations. Our proposed framework is composed of modified networks for extracting the most correlated intermediate representations from off-pose thermal and frontal visible face imagery, a sub-network to jointly bridge domain and pose gaps, and a joint-loss function comprised of cross-spectrum and pose-correction losses. We demonstrate efficacy and advantages of the proposed method by evaluating on three thermal-visible datasets: ARL Visible-to-Thermal Face, ARL Multimodal Face, and Tufts Face. Although DPIF focuses on learning to match off-pose thermal to frontal visible faces, we also show that DPIF enhances performance when matching frontal thermal face images to frontal visible face images.