While the researches on single image super-resolution (SISR), especially equipped with deep neural networks (DNNs), have achieved tremendous successes recently, they still suffer from two major limitations. Firstly, the real image degradation is usually unknown and highly variant from one to another, making it extremely hard to train a single model to handle the general SISR task. Secondly, most of current methods mainly focus on the downsampling process of the degradation, but ignore or underestimate the inevitable noise contamination. For example, the commonly-used independent and identically distributed (i.i.d.) Gaussian noise distribution always largely deviates from the real image noise (e.g., camera sensor noise), which limits their performance in real scenarios. To address these issues, this paper proposes a model-based unsupervised SISR method to deal with the general SISR task with unknown degradations. Instead of the traditional i.i.d. Gaussian noise assumption, a novel patch-based non-i.i.d. noise modeling method is proposed to fit the complex real noise. Besides, a deep generator parameterized by a DNN is used to map the latent variable to the high-resolution image, and the conventional hyper-Laplacian prior is also elaborately embedded into such generator to further constrain the image gradients. Finally, a Monte Carlo EM algorithm is designed to solve our model, which provides a general inference framework to update the image generator both w.r.t. the latent variable and the network parameters. Comprehensive experiments demonstrate that the proposed method can evidently surpass the current state of the art (SotA) method (about 1dB PSNR) not only with a slighter model (0.34M vs. 2.40M) but also faster speed.
Scene text image super-resolution (STISR) aims to improve the resolution and visual quality of low-resolution (LR) scene text images, and consequently boost the performance of text recognition. However, most of existing STISR methods regard text images as natural scene images, ignoring the categorical information of text. In this paper, we make an inspiring attempt to embed categorical text prior into STISR model training. Specifically, we adopt the character probability sequence as the text prior, which can be obtained conveniently from a text recognition model. The text prior provides categorical guidance to recover high-resolution (HR) text images. On the other hand, the reconstructed HR image can refine the text prior in return. Finally, we present a multi-stage text prior guided super-resolution (TPGSR) framework for STISR. Our experiments on the benchmark TextZoom dataset show that TPGSR can not only effectively improve the visual quality of scene text images, but also significantly improve the text recognition accuracy over existing STISR methods. Our model trained on TextZoom also demonstrates certain generalization capability to the LR images in other datasets.
Faceted summarization provides briefings of a document from different perspectives. Readers can quickly comprehend the main points of a long document with the help of a structured outline. However, little research has been conducted on this subject, partially due to the lack of large-scale faceted summarization datasets. In this study, we present FacetSum, a faceted summarization benchmark built on Emerald journal articles, covering a diverse range of domains. Different from traditional document-summary pairs, FacetSum provides multiple summaries, each targeted at specific sections of a long document, including the purpose, method, findings, and value. Analyses and empirical results on our dataset reveal the importance of bringing structure into summaries. We believe FacetSum will spur further advances in summarization research and foster the development of NLP systems that can leverage the structured information in both long texts and summaries.
Domain adaptation becomes more challenging with increasing gaps between source and target domains. Motivated from an empirical analysis on the reliability of labeled source data for the use of distancing target domains, we propose self-training of auxiliary models (AuxSelfTrain) that learns models for intermediate domains and gradually combats the distancing shifts across domains. We introduce evolving intermediate domains as combinations of decreasing proportion of source data and increasing proportion of target data, which are sampled to minimize the domain distance between consecutive domains. Then the source model could be gradually adapted for the use in the target domain by self-training of auxiliary models on evolving intermediate domains. We also introduce an enhanced indicator for sample selection via implicit ensemble and extend the proposed method to semi-supervised domain adaptation. Experiments on benchmark datasets of unsupervised and semi-supervised domain adaptation verify its efficacy.
The complex nature of combining localization and classification in object detection has resulted in the flourished development of methods. Previous works tried to improve the performance in various object detection heads but failed to present a unified view. In this paper, we present a novel dynamic head framework to unify object detection heads with attentions. By coherently combining multiple self-attention mechanisms between feature levels for scale-awareness, among spatial locations for spatial-awareness, and within output channels for task-awareness, the proposed approach significantly improves the representation ability of object detection heads without any computational overhead. Further experiments demonstrate that the effectiveness and efficiency of the proposed dynamic head on the COCO benchmark. With a standard ResNeXt-101-DCN backbone, we largely improve the performance over popular object detectors and achieve a new state-of-the-art at 54.0 AP. Furthermore, with latest transformer backbone and extra data, we can push current best COCO result to a new record at 60.6 AP. The code will be released at https://github.com/microsoft/DynamicHead.
Benefit from large-scale training data, recent advances in Siamese-based object tracking have achieved compelling results on the normal sequences. Whilst Siamese-based trackers assume training and test data follow an identical distribution. Suppose there is a set of foggy or rainy test sequences, it cannot be guaranteed that the trackers trained on the normal images perform well on the data belonging to other domains. The problem of domain shift among training and test data has already been discussed in object detection and semantic segmentation areas, which, however, has not been investigated for visual tracking. To this end, based on SiamRPN++, we introduce a Domain Adaptive SiamRPN++, namely DASiamRPN++, to improve the cross-domain transferability and robustness of a tracker. Inspired by A-distance theory, we present two domain adaptive modules, Pixel Domain Adaptation (PDA) and Semantic Domain Adaptation (SDA). The PDA module aligns the feature maps of template and search region images to eliminate the pixel-level domain shift caused by weather, illumination, etc. The SDA module aligns the feature representations of the tracking target's appearance to eliminate the semantic-level domain shift. PDA and SDA modules reduce the domain disparity by learning domain classifiers in an adversarial training manner. The domain classifiers enforce the network to learn domain-invariant feature representations. Extensive experiments are performed on the standard datasets of two different domains, including synthetic foggy and TIR sequences, which demonstrate the transferability and domain adaptability of the proposed tracker.
Vision transformers (ViTs) have recently received explosive popularity, but their enormous model sizes and training costs remain daunting. Conventional post-training pruning often incurs higher training budgets. In contrast, this paper aims to trim down both the training memory overhead and the inference complexity, without sacrificing the achievable accuracy. We launch and report the first-of-its-kind comprehensive exploration, on taking a unified approach of integrating sparsity in ViTs "from end to end". Specifically, instead of training full ViTs, we dynamically extract and train sparse subnetworks, while sticking to a fixed small parameter budget. Our approach jointly optimizes model parameters and explores connectivity throughout training, ending up with one sparse network as the final output. The approach is seamlessly extended from unstructured to structured sparsity, the latter by considering to guide the prune-and-grow of self-attention heads inside ViTs. For additional efficiency gains, we further co-explore data and architecture sparsity, by plugging in a novel learnable token selector to adaptively determine the currently most vital patches. Extensive results on ImageNet with diverse ViT backbones validate the effectiveness of our proposals which obtain significantly reduced computational cost and almost unimpaired generalization. Perhaps most surprisingly, we find that the proposed sparse (co-)training can even improve the ViT accuracy rather than compromising it, making sparsity a tantalizing "free lunch". For example, our sparsified DeiT-Small at (5%, 50%) sparsity for (data, architecture), improves 0.28% top-1 accuracy, and meanwhile enjoys 49.32% FLOPs and 4.40% running time savings. Our codes are available at https://github.com/VITA-Group/SViTE.
Vision transformers (ViTs) have recently received explosive popularity, but their enormous model sizes and training costs remain daunting. Conventional post-training pruning often incurs higher training budgets. In contrast, this paper aims to trim down both the training memory overhead and the inference complexity, without scarifying the achievable accuracy. We launch and report the first-of-its-kind comprehensive exploration, on taking a unified approach of integrating sparsity in ViTs "from end to end". Specifically, instead of training full ViTs, we dynamically extract and train sparse subnetworks, while sticking to a fixed small parameter budget. Our approach jointly optimizes model parameters and explores connectivity throughout training, ending up with one sparse network as the final output. The approach is seamlessly extended from unstructured to structured sparsity, the latter by considering to guide the prune-and-grow of self-attention heads inside ViTs. For additional efficiency gains, we further co-explore data and architecture sparsity, by plugging in a novel learnable token selector to adaptively determine the currently most vital patches. Extensive results validate the effectiveness of our proposals on ImageNet with diverse ViT backbones. For instance, at 40% structured sparsity, our sparsified DeiT-Base can achieve 0.42% accuracy gain, at 33.13% and 24.70% running time} savings, compared to its dense counterpart. Perhaps most surprisingly, we find that the proposed sparse (co-)training can even improve the ViT accuracy rather than compromising it, making sparsity a tantalizing "free lunch". For example, our sparsified DeiT-Small at 5%, 50% sparsity for (data, architecture), improves 0.28% top-1 accuracy and meanwhile enjoys 49.32% FLOPs and 4.40% running time savings.
Unsupervised domain adaptation (UDA) and semi-supervised learning (SSL) are two typical strategies to reduce expensive manual annotations in machine learning. In order to learn effective models for a target task, UDA utilizes the available labeled source data, which may have different distributions from unlabeled samples in the target domain, while SSL employs few manually annotated target samples. Although UDA and SSL are seemingly very different strategies, we find that they are closely related in terms of task objectives and solutions, and SSL is a special case of UDA problems. Based on this finding, we further investigate whether SSL methods work on UDA tasks. By adapting eight representative SSL algorithms on UDA benchmarks, we show that SSL methods are strong UDA learners. Especially, state-of-the-art SSL methods significantly outperform existing UDA methods on the challenging UDA benchmark of DomainNet, and state-of-the-art UDA methods could be further enhanced with SSL techniques. We thus promote that SSL methods should be employed as baselines in future UDA studies and expect that the revealed relationship between UDA and SSL could shed light on future UDA development. Codes are available at \url{https://github.com/YBZh}.