Recent Anomaly Detection (AD) methods have achieved great success with In-Distribution (ID) data. However, real-world data often exhibits distribution shift, causing huge performance decay on traditional AD methods. From this perspective, few previous work has explored AD with distribution shift, and the distribution-invariant normality learning has been proposed based on the Reverse Distillation (RD) framework. However, we observe the misalignment issue between the teacher and the student network that causes detection failure, thereby propose FiCo, Filter or Compensate, to address the distribution shift issue in AD. FiCo firstly compensates the distribution-specific information to reduce the misalignment between the teacher and student network via the Distribution-Specific Compensation (DiSCo) module, and secondly filters all abnormal information to capture distribution-invariant normality with the Distribution-Invariant Filter (DiIFi) module. Extensive experiments on three different AD benchmarks demonstrate the effectiveness of FiCo, which outperforms all existing state-of-the-art (SOTA) methods, and even achieves better results on the ID scenario compared with RD-based methods.

Composed Image Retrieval (CIR) is a challenging vision-language (VL) task that retrieves target images using multi-modal (image+text) queries. Although significant progress has been made in existing CIR approaches, their deployment in resource-constrained scenarios remains problematic. To address this issue, we propose a novel framework, named Slot Inversion for Asymmetric Composed Image Retrieval (Slot4ACir), where an asymmetric retrieval scheme is adopted: lightweight models are employed on the query side, while large-scale VL models operate on the gallery side. Specifically, we introduce a lightweight inversion module based on slot attention, which maps an image into multiple textual tokens with distinct semantics. Additionally, the LLM sampler is proposed to facilitate richer semantic interactions, and the distillation alignment (DTA) loss enables the extraction of more informative representations. Extensive experiments on two popular benchmarks demonstrate the effectiveness of our approach. The code is now available at https://github.com/JThuge/Slot4ACir.

Domain Generalization (DG) aims to resolve distribution shifts between source and target domains, and current DG methods are default to the setting that data from source and target domains share identical categories. Nevertheless, there exists unseen classes from target domains in practical scenarios. To address this issue, Open Set Domain Generalization (OSDG) has emerged and several methods have been exclusively proposed. However, most existing methods adopt complex architectures with slight improvement compared with DG methods. Recently, vision-language models (VLMs) have been introduced in DG following the fine-tuning paradigm, but consume huge training overhead with large vision models. Therefore, in this paper, we innovate to transfer knowledge from VLMs to lightweight vision models and improve the robustness by introducing Perturbation Distillation (PD) from three perspectives, including Score, Class and Instance (SCI), named SCI-PD. Moreover, previous methods are oriented by the benchmarks with identical and fixed splits, ignoring the divergence between source domains. These methods are revealed to suffer from sharp performance decay with our proposed new benchmark Hybrid Domain Generalization (HDG) and a novel metric H2-CV, which construct various splits to comprehensively assess the robustness of algorithms. Extensive experiments demonstrate that our method outperforms state-of-the-art algorithms on multiple datasets, especially improving the robustness when confronting data scarcity.

Deep learning methods usually succeed in independent and identically distributed (IID) data distribution, but suffer from sharp performance decay in real-world out-of-distribution (OOD) data. OOD generalization emerges to alleviate the large distribution shift between source and target domains. Recently, domain-invariant learning has boosted the research on OOD generalization, but most methods indulge complex architectures and training strategies. Hence, we propose a simple yet effective Selective Cross-Correlation Consistency (SC3) loss to align the cross-correlation matrix of features from identical categories. Specifically, we design the Semantic-Oriented Selection (SOS) algorithm in SC3loss to eliminate negative effects on spurious channels. Extensive experiments demonstrate that the SC3loss achieves superior performance on multiple OOD scenarios, including domain generalization (DG) and single domain gener-alization (SDG) tasks. Also, our loss consumes negligible computational resource which conforms to real-world applications. Source code is available at https://github.com/znchen666/SC3.

Domain Generalization (DG) aims to develop models that can learn from data in source domains and generalize to unseen target domains. Recently, some domain generalization algorithms have emerged, but most of them were designed with complex modules. Among all the prior methods under DG settings, contrastive learning has become a promising solution for simplicity and efficiency. However, existing contrastive learning neglects distribution shifts that causes severe domain confusions. In this paper, we propose an instance paradigm contrastive learning framework, introducing contrast between original features and novel paradigms to alleviate domain-specific distractions. And then we explore hard-pair information, an essential factor in contrastive learning, based on domain label and feature similarity. Moreover, to produce domain-invariant instance paradigms, we generate multiple views of the original images and design a novel channel-wise attention mechanism to dynamically combine features from all the views. Furthermore, a test-time feature integration module is designed to mimic the paradigms during the training process to improve generalization ability. Extensive experiments show that our method achieves state-of-the-art performance. The proposed algorithm can also serve as a plug-and-play module which improves performance of existing methods with a relatively large margin.

Person re-identification (ReID) has achieved great improvement under supervised settings, but suffers from considerable degradation when large distribution shifts between training and testing sets exist. Domain generalization (DG ReID) emerges to promote the generalization ability of models, overcoming the distribution shifts issue between source domains and unseen target domains. Among most prior methods in DG ReID, instance normalization (IN) serves as a promising solution for removing domain-specific information, however, it damages the discriminative ability simultaneously. In this article, we propose a new normalization method called Cluster-Instance Normalization (CINorm) to extract information from clusters for information compensation. The relations between samples in a batch can be mined to establish evolving clusters with aggregated samples during the forward training process. In this way, high intra-cluster congregation can eliminate the impacts of outliers to avoid overfitting, and high inter-cluster variances can synthesize diverse novel statistics to compensate discriminative information. Therefore, a Relation-Aware Normalization (RANorm) with a Dynamic ReCalibration (DRC) module is designed to integrate normalized features between evolving clusters and instances efficiently. Furthermore, a novel Group-based Triplet (G-Triplet) loss is proposed to divide a batch into multiple groups with greater compactness for hard-pair mining. Extensive experiments show that our method outperforms state-of-the-art algorithms on multiple DG benchmarks by a large margin. The proposed method can also achieve superior performance on image classification tasks under DG settings without using domain labels.