Transformers have shown remarkable performance, however, their architecture design is a time-consuming process that demands expertise and trial-and-error. Thus, it is worthwhile to investigate efficient methods for automatically searching high-performance Transformers via Transformer Architecture Search (TAS). In order to improve the search efficiency, training-free proxy based methods have been widely adopted in Neural Architecture Search (NAS). Whereas, these proxies have been found to be inadequate in generalizing well to Transformer search spaces, as confirmed by several studies and our own experiments. This paper presents an effective scheme for TAS called TRansformer Architecture search with ZerO-cost pRoxy guided evolution (T-Razor) that achieves exceptional efficiency. First, through theoretical analysis, we discover that the synaptic diversity of multi-head self-attention (MSA) and the saliency of multi-layer perceptron (MLP) are correlated with the performance of corresponding Transformers. The properties of synaptic diversity and synaptic saliency motivate us to introduce the ranks of synaptic diversity and saliency that denoted as DSS++ for evaluating and ranking Transformers. DSS++ incorporates correlation information among sampled Transformers to provide unified scores for both synaptic diversity and synaptic saliency. We then propose a block-wise evolution search guided by DSS++ to find optimal Transformers. DSS++ determines the positions for mutation and crossover, enhancing the exploration ability. Experimental results demonstrate that our T-Razor performs competitively against the state-of-the-art manually or automatically designed Transformer architectures across four popular Transformer search spaces. Significantly, T-Razor improves the searching efficiency across different Transformer search spaces, e.g., reducing required GPU days from more than 24 to less than 0.4 and outperforming existing zero-cost approaches. We also apply T-Razor to the BERT search space and find that the searched Transformers achieve competitive GLUE results on several Neural Language Processing (NLP) datasets. This work provides insights into training-free TAS, revealing the usefulness of evaluating Transformers based on the properties of their different blocks.

Vision transformers (ViTs) have recently received explosive popularity, but the huge computational cost is still a severe issue. Since the computation complexity of ViT is quadratic with respect to the input sequence length, a mainstream paradigm for computation reduction is to reduce the number of tokens. Existing designs include structured spatial compression that uses a progressive shrinking pyramid to reduce the computations of large feature maps, and unstructured token pruning that dynamically drops redundant tokens. However, the limitation of existing token pruning lies in two folds: 1) the incomplete spatial structure caused by pruning is not compatible with structured spatial compression that is commonly used in modern deep-narrow transformers; 2) it usually requires a time-consuming pre-training procedure. To tackle the limitations and expand the applicable scenario of token pruning, we present Evo-ViT, a self-motivated slow-fast token evolution approach for vision transformers. Specifically, we conduct unstructured instance-wise token selection by taking advantage of the simple and effective global class attention that is native to vision transformers. Then, we propose to update the selected informative tokens and uninformative tokens with different computation paths, namely, slow-fast updating. Since slow-fast updating mechanism maintains the spatial structure and information flow, Evo-ViT can accelerate vanilla transformers of both flat and deep-narrow structures from the very beginning of the training process. Experimental results demonstrate that our method significantly reduces the computational cost of vision transformers while maintaining comparable performance on image classification. For example, our method accelerates DeiT-S by over 60% throughput while only sacrificing 0.4% top-1 accuracy on ImageNet-1K, outperforming current token pruning methods on both accuracy and efficiency.

Recently, Vision Transformer (ViT) has achieved remarkable success in several computer vision tasks. The progresses are highly relevant to the architecture design, then it is worthwhile to propose Transformer Architecture Search (TAS) to search for better ViTs automatically. However, current TAS methods are time-consuming and existing zero-cost proxies in CNN do not generalize well to the ViT search space according to our experimental observations. In this paper, for the first time, we investigate how to conduct TAS in a training-free manner and devise an effective training-free TAS (TF-TAS) scheme. Firstly, we observe that the properties of multi-head self-attention (MSA) and multi-layer perceptron (MLP) in ViTs are quite different and that the synaptic diversity of MSA affects the performance notably. Secondly, based on the observation, we devise a modular strategy in TF-TAS that evaluates and ranks ViT architectures from two theoretical perspectives: synaptic diversity and synaptic saliency, termed as DSS-indicator. With DSS-indicator, evaluation results are strongly corre-lated with the test accuracies of ViT models. Experimental results demonstrate that our TF- TAS achieves a competitive performance against the state-of-the-art manually or automatically design ViT architectures, and it promotes the searching efficiency in ViT search space greatly: from about 24 GPU days to less than 0.5 GPU days. Moreover, the proposed DSS-indicator outperforms the existing cutting-edge zero-cost approaches (e.g., TE-score and NASWOT).

Transformers, the dominant architecture for natural language processing, have also recently attracted much attention from computational visual media researchers due to their capacity for long-range representation and high performance. Transformers are sequence-to-sequence models, which use a self-attention mechanism rather than the RNN sequential structure. Thus, such models can be trained in parallel and can represent global information. This study comprehensively surveys recent visual transformer works. We categorize them according to task scenario: backbone design, high-level vision, low-level vision and generation, and multimodal learning. Their key ideas are also analyzed. Differing from previous surveys, we mainly focus on visual transformer methods in low-level vision and generation. The latest works on backbone design are also reviewed in detail. For ease of understanding, we precisely describe the main contributions of the latest works in the form of tables. As well as giving quantitative comparisons, we also present image results for low-level vision and generation tasks. Computational costs and source code links for various important works are also given in this survey to assist further development.

Great progress has been achieved in domain adaptation in decades. Existing works are always based on an ideal assumption that testing target domains are independent and identically distributed with training target domains. However, due to unpredictable corruptions (e.g., noise and blur) in real data, such as web images and real-world object detection, domain adaptation methods are increasingly required to be corruption robust on target domains. We investigate a new task, corruption-agnostic robust domain adaptation (CRDA), to be accurate on original data and robust against unavailable-for-training corruptions on target domains. This task is non-trivial due to the large domain discrepancy and unsupervised target domains. We observe that simple combinations of popular methods of domain adaptation and corruption robustness have suboptimal CRDA results. We propose a new approach based on two technical insights into CRDA, as follows: (1) an easy-to-plug module called domain discrepancy generator (DDG) that generates samples that enlarge domain discrepancy to mimic unpredictable corruptions; (2) a simple but effective teacher-student scheme with contrastive loss to enhance the constraints on target domains. Experiments verify that DDG maintains or even improves its performance on original data and achieves better corruption robustness than baselines. Our code is available at: https://github.com/YifanXu74/CRDA .

Face anti-spoofing approach based on domain generalization (DG) has drawn growing attention due to its robustness for unseen scenarios. Existing DG methods assume that the domain label is known. However, in real-world applications, the collected dataset always contains mixture domains, where the domain label is unknown. In this case, most of existing methods may not work. Further, even if we can obtain the domain label as existing methods, we think this is just a sub-optimal partition. To overcome the limitation, we propose domain dynamic adjustment meta-learning (D$^2$AM) without using domain labels, which iteratively divides mixture domains via discriminative domain representation and trains a generalizable face anti-spoofing with meta-learning. Specifically, we design a domain feature based on Instance Normalization (IN) and propose a domain representation learning module (DRLM) to extract discriminative domain features for clustering. Moreover, to reduce the side effect of outliers on clustering performance, we additionally utilize maximum mean discrepancy (MMD) to align the distribution of sample features to a prior distribution, which improves the reliability of clustering. Extensive experiments show that the proposed method outperforms conventional DG-based face anti-spoofing methods, including those utilizing domain labels. Furthermore, we enhance the interpretability through visualization.

Face anti-spoofing approaches based on domain generalization (DG) have drawn growing attention due to their robustness for unseen scenarios. Previous methods treat each sample from multiple domains indiscriminately during the training process, and endeavor to extract a common feature space to improve the generalization. However, due to complex and biased data distribution, directly treating them equally will corrupt the generalization ability. To settle the issue, we propose a novel Dual Reweighting Domain Generalization (DRDG) framework which iteratively reweights the relative importance between samples to further improve the generalization. Concretely, Sample Reweighting Module is first proposed to identify samples with relatively large domain bias, and reduce their impact on the overall optimization. Afterwards, Feature Reweighting Module is introduced to focus on these samples and extract more domain-irrelevant features via a self-distilling mechanism. Combined with the domain discriminator, the iteration of the two modules promotes the extraction of generalized features. Extensive experiments and visualizations are presented to demonstrate the effectiveness and interpretability of our method against the state-of-the-art competitors.

Distinguishing aesthetically pleasing food photos from others is an important visual analysis task for social media and ranking systems related to food. Nevertheless, aesthetic assessment of food images remains a challenging and relatively unexplored task, largely due to the lack of related food image datasets and practical knowledge. Thus, we present the Gourmet Photography Dataset (GPD), the first large-scale dataset for aesthetic assessment of food photos. It contains 24,000 images with corresponding binary aesthetic labels, covering a large variety of foods and scenes. We also provide a non-stationary regularization method to combat over-fitting and enhance the ability of tuned models to generalize. Quantitative results from extensive experiments, including a generalization ability test, verify that neural networks trained on the GPD achieve comparable performance to human experts on the task of aesthetic assessment. We reveal several valuable findings to support further research and applications related to visual aesthetic analysis of food images. To encourage further research, we have made the GPD publicly available at https://github.com/Openning07/GPA .

Visual aesthetic assessment has been an active research field for decades. Although latest methods have achieved promising performance on benchmark datasets, they typically rely on a large number of manual annotations including both aesthetic labels and related image attributes. In this paper, we revisit the problem of image aesthetic assessment from the self-supervised feature learning perspective. Our motivation is that a suitable feature representation for image aesthetic assessment should be able to distinguish different expert-designed image manipulations, which have close relationships with negative aesthetic effects. To this end, we design two novel pretext tasks to identify the types and parameters of editing operations applied to synthetic instances. The features from our pretext tasks are then adapted for a one-layer linear classifier to evaluate the performance in terms of binary aesthetic classification. We conduct extensive quantitative experiments on three benchmark datasets and demonstrate that our approach can faithfully extract aesthetics-aware features and outperform alternative pretext schemes. Moreover, we achieve comparable results to state-of-the-art supervised methods that use 10 million labels from ImageNet.

Object detection has achieved remarkable progress in the past decade. However, the detection of oriented and densely packed objects remains challenging because of following inherent reasons: (1) receptive fields of neurons are all axis-aligned and of the same shape, whereas objects are usually of diverse shapes and align along various directions; (2) detection models are typically trained with generic knowledge and may not generalize well to handle specific objects at test time; (3) the limited dataset hinders the development on this task. To resolve the first two issues, we present a dynamic refinement network that consists of two novel components, i.e., a feature selection module (FSM) and a dynamic refinement head (DRH). Our FSM enables neurons to adjust receptive fields in accordance with the shapes and orientations of target objects, whereas the DRH empowers our model to refine the prediction dynamically in an object-aware manner. To address the limited availability of related benchmarks, we collect an extensive and fully annotated dataset, namely, SKU110K-R, which is relabeled with oriented bounding boxes based on SKU110K. We perform quantitative evaluations on several publicly available benchmarks including DOTA, HRSC2016, SKU110K, and our own SKU110K-R dataset. Experimental results show that our method achieves consistent and substantial gains compared with baseline approaches. Our source code and dataset will be released to encourage follow-up research.

Direct regression and anchor are the two mainly effective and prevailing mechanisms in the paradigm of scene text detection. However, the use of direct regression-based methods may be challenging during optimization without the help of anchors as references. Unfortunately, the anchor-based methods always suffer from the careful design of the anchors, degrading the robustness to complex scenes. To address the above-mentioned problems, we propose a novel hidden anchor mechanism (HAM) especially for scene text detection. The predictions of anchors are innovatively regarded as hidden layers, and the weighted sum of the predictions is integrated into a direct regression-based network. Hence, the architecture of our HAM still has the characteristic of simplicity as with direct regression-based methods. Moreover, it is easier to optimize anchors as references with this type of method than with direct regression-based methods. In this way, our network can take advantage of both direct regression and anchor mechanisms. In addition, we decouple three kinds of one-dimensional anchors from three-dimensional anchors, greatly reducing the number of anchors in text bounding box matching without performance degradation. We also propose a post-processing technique for long text detection, named iterative regression box (IRB), which takes a few additional computational costs and can be easily generalized to other methods. Experiments on several public datasets demonstrate that the proposed method achieves state-of-the-art performance. Code is available athttps://github.com/hjbplayer/HAM.

Aggregation structures with explicit information, such as image attributes and scene semantics, are effective and popular for intelligent systems for assessing aesthetics of visual data. However, useful information may not be available due to the high cost of manual annotation and expert design. In this paper, we present a novel multi-patch (MP) aggregation method for image aesthetic assessment. Different from state-of-the-art methods, which augment an MP aggregation network with various visual attributes, we train the model in an end-to-end manner with aesthetic labels only (i.e., aesthetically positive or negative). We achieve the goal by resorting to an attention-based mechanism that adaptively adjusts the weight of each patch during the training process to improve learning efficiency. In addition, we propose a set of objectives with three typical attention mechanisms (i.e., average, minimum, and adaptive) and evaluate their effectiveness on the Aesthetic Visual Analysis (AVA) benchmark. Numerical results show that our approach outperforms existing methods by a large margin. We further verify the effectiveness of the proposed attention-based objectives via ablation studies and shed light on the design of aesthetic assessment systems.

The study of face alignment has been an area of intense research in computer vision, with its achievements widely used in computer graphics applications. The performance of various face alignment methods is often image-dependent or somewhat random because of their own strategy. This study aims to develop a method that can select an input image with good face alignment results from many results produced by a single method or multiple ones. The task is challenging because different face alignment results need to be evaluated without any ground truth. This study addresses this problem by designing a feasible feature extraction scheme to measure the quality of face alignment results. The feature is then used in various machine learning algorithms to rank different face alignment results. Our experiments show that our method is promising for ranking face alignment results and is able to pick good face alignment results, which can enhance the overall performance of a face alignment method with a random strategy. We demonstrate the usefulness of our ranking-enhanced face alignment algorithm in two practical applications: face cartoon stylization and digital face makeup.