Masked image modeling (MIM), an emerging self-supervised pre-training method, has shown impressive success across numerous downstream vision tasks with Vision transformers (ViT). Its underlying idea is simple: a portion of the input image is randomly masked out and then reconstructed via the pre-text task. However, why MIM works well is not well explained, and previous studies insist that MIM primarily works for the Transformer family but is incompatible with CNNs. In this paper, we first study interactions among patches to understand what knowledge is learned and how it is acquired via the MIM task. We observe that MIM essentially teaches the model to learn better middle-level interactions among patches and extract more generalized features. Based on this fact, we propose an Architecture-Agnostic Masked Image Modeling framework (A$^2$MIM), which is compatible with not only Transformers but also CNNs in a unified way. Extensive experiments on popular benchmarks show that our A$^2$MIM learns better representations and endows the backbone model with the stronger capability to transfer to various downstream tasks for both Transformers and CNNs.
Zero-shot learning (ZSL) aims to recognize novel classes by transferring semantic knowledge from seen classes to unseen ones. Semantic knowledge is learned from attribute descriptions shared between different classes, which act as strong priors for localizing object attributes that represent discriminative region features, enabling significant visual-semantic interaction. Although some attention-based models have attempted to learn such region features in a single image, the transferability and discriminative attribute localization of visual features are typically neglected. In this paper, we propose an attribute-guided Transformer network, termed TransZero, to refine visual features and learn attribute localization for discriminative visual embedding representations in ZSL. Specifically, TransZero takes a feature augmentation encoder to alleviate the cross-dataset bias between ImageNet and ZSL benchmarks, and improves the transferability of visual features by reducing the entangled relative geometry relationships among region features. To learn locality-augmented visual features, TransZero employs a visual-semantic decoder to localize the image regions most relevant to each attribute in a given image, under the guidance of semantic attribute information. Then, the locality-augmented visual features and semantic vectors are used to conduct effective visual-semantic interaction in a visual-semantic embedding network. Extensive experiments show that TransZero achieves the new state of the art on three ZSL benchmarks. The codes are available at: \url{https://github.com/shiming-chen/TransZero}.
Zero-shot learning (ZSL) tackles the unseen class recognition problem, transferring semantic knowledge from seen classes to unseen ones. Typically, to guarantee desirable knowledge transfer, a common (latent) space is adopted for associating the visual and semantic domains in ZSL. However, existing common space learning methods align the semantic and visual domains by merely mitigating distribution disagreement through one-step adaptation. This strategy is usually ineffective due to the heterogeneous nature of the feature representations in the two domains, which intrinsically contain both distribution and structure variations. To address this and advance ZSL, we propose a novel hierarchical semantic-visual adaptation (HSVA) framework. Specifically, HSVA aligns the semantic and visual domains by adopting a hierarchical two-step adaptation, i.e., structure adaptation and distribution adaptation. In the structure adaptation step, we take two task-specific encoders to encode the source data (visual domain) and the target data (semantic domain) into a structure-aligned common space. To this end, a supervised adversarial discrepancy (SAD) module is proposed to adversarially minimize the discrepancy between the predictions of two task-specific classifiers, thus making the visual and semantic feature manifolds more closely aligned. In the distribution adaptation step, we directly minimize the Wasserstein distance between the latent multivariate Gaussian distributions to align the visual and semantic distributions using a common encoder. Finally, the structure and distribution adaptation are derived in a unified framework under two partially-aligned variational autoencoders. Extensive experiments on four benchmark datasets demonstrate that HSVA achieves superior performance on both conventional and generalized ZSL. The code is available at \url{https://github.com/shiming-chen/HSVA} .
Self-supervised Multi-view stereo (MVS) with a pretext task of image reconstruction has achieved significant progress recently. However, previous methods are built upon intuitions, lacking comprehensive explanations about the effectiveness of the pretext task in self-supervised MVS. To this end, we propose to estimate epistemic uncertainty in self-supervised MVS, accounting for what the model ignores. Specially, the limitations can be categorized into two types: ambiguious supervision in foreground and invalid supervision in background. To address these issues, we propose a novel Uncertainty reduction Multi-view Stereo (UMVS) framework for self-supervised learning. To alleviate ambiguous supervision in foreground, we involve extra correspondence prior with a flow-depth consistency loss. The dense 2D correspondence of optical flows is used to regularize the 3D stereo correspondence in MVS. To handle the invalid supervision in background, we use Monte-Carlo Dropout to acquire the uncertainty map and further filter the unreliable supervision signals on invalid regions. Extensive experiments on DTU and Tank&Temples benchmark show that our U-MVS framework achieves the best performance among unsupervised MVS methods, with competitive performance with its supervised opponents.
While semi-supervised learning (SSL) has received tremendous attentions in many machine learning tasks due to its successful use of unlabeled data, existing SSL algorithms use either all unlabeled examples or the unlabeled examples with a fixed high-confidence prediction during the training progress. However, it is possible that too many correct/wrong pseudo labeled examples are eliminated/selected. In this work we develop a simple yet powerful framework, whose key idea is to select a subset of training examples from the unlabeled data when performing existing SSL methods so that only the unlabeled examples with pseudo labels related to the labeled data will be used to train models. The selection is performed at each updating iteration by only keeping the examples whose losses are smaller than a given threshold that is dynamically adjusted through the iteration. Our proposed approach, Dash, enjoys its adaptivity in terms of unlabeled data selection and its theoretical guarantee. Specifically, we theoretically establish the convergence rate of Dash from the view of non-convex optimization. Finally, we empirically demonstrate the effectiveness of the proposed method in comparison with state-of-the-art over benchmarks.
Self-supervised contrastive learning has demonstrated great potential in learning visual representations. Despite their success on various downstream tasks such as image classification and object detection, self-supervised pre-training for fine-grained scenarios is not fully explored. In this paper, we first point out that current contrastive methods are prone to memorizing background/foreground texture and therefore have a limitation in localizing the foreground object. Analysis suggests that learning to extract discriminative texture information and localization are equally crucial for self-supervised pre-training under fine-grained scenarios. Based on our findings, we introduce Cross-view Saliency Alignment (CVSA), a contrastive learning framework that first crops and swaps saliency regions of images as a novel view generation and then guides the model to localize on the foreground object via a cross-view alignment loss. Extensive experiments on four popular fine-grained classification benchmarks show that CVSA significantly improves the learned representation.
Face recognition has achieved significant progress in deep-learning era due to the ultra-large-scale and well-labeled datasets. However, training on ultra-large-scale datasets is time-consuming and takes up a lot of hardware resource. Therefore, designing an efficient training approach is crucial and indispensable. The heavy computational and memory costs mainly result from the high dimensionality of the Fully-Connected (FC) layer. Specifically, the dimensionality is determined by the number of face identities, which can be million-level or even more. To this end, we propose a novel training approach for ultra-large-scale face datasets, termed Faster Face Classification (F$^2$C). In F$^2$C, we first define a Gallery Net and a Probe Net that are used to generate identities' centers and extract faces' features for face recognition, respectively. Gallery Net has the same structure as Probe Net and inherits the parameters from Probe Net with a moving average paradigm. After that, to reduce the training time and hardware costs of the FC layer, we propose a Dynamic Class Pool (DCP) that stores the features from Gallery Net and calculates the inner product (logits) with positive samples (whose identities are in the DCP) in each mini-batch. DCP can be regarded as a substitute for the FC layer but it is far smaller, thus greatly reducing the computational and memory costs. For negative samples (whose identities are not in DCP), we minimize the cosine similarities between negative samples and those in DCP. Then, to improve the update efficiency of DCP's parameters, we design a dual data-loader including identity-based and instance-based loaders to generate a certain of identities and samples in mini-batches.