Logical anomaly detection aims to identify inconsistencies in object relationships and scene semantics, which is a challenging task requiring high-level reasoning capabilities that go beyond low-level structural cues. Existing methods rely heavily on large amounts of labeled data, struggle in zero-shot scenarios and lack deep semantic understanding in complex scenes with multiple objects. To address this, we propose VLLM-LAD, a visual large language model framework specifically designed for zero-shot logical anomaly detection. It incorporates a Local Detail Visual Perceiver (LDVP) module for fine-grained anomaly localization and a Global Semantic Feature Fusion (GSFF) module for enhanced multimodal reasoning. We also introduce a global-local-regional contrastive learning strategy to align visual and textual representations, thereby achieving robust zero-shot performance. Extensive experiments on MVTec LOCO demonstrate that VLLM-LAD outperforms existing methods in accuracy and interpretability. To further advance research in this area, we introduce the RLAD dataset. This large-scale, diverse dataset is designed for logical anomaly detection, encompassing complex scenarios that demand reasoning about object relationships and global semantics. Our method not only excels on the RLAD dataset but also provides detailed textual explanations and segmentation masks, enhancing its transparency and applicability to real-world tasks. This work establishes a new benchmark for zero-shot logical anomaly detection, offering a robust framework and a valuable dataset to propel future research.

Most of existing category-level object pose estimation methods devote to learning the object category information from point cloud modality. However, the scale of 3D datasets is limited due to the high cost of 3D data collection and annotation. Consequently, the category features extracted from these limited point cloud samples may not be comprehensive. This motivates us to investigate whether we can draw on knowledge of other modalities to obtain category information. Inspired by this motivation, we propose CLIPose, a novel 6D pose framework that employs the pre-trained vision-language model to develop better learning of object category information, which can fully leverage abundant semantic knowledge in image and text modalities. To make the 3D encoder learn category-specific features more efficiently, we align representations of three modalities in feature space via multi-modal contrastive learning. In addition to exploiting the pre-trained knowledge of the CLIP’s model, we also expect it to be more sensitive with pose parameters. Therefore, we introduce a prompt tuning approach to fine-tune image encoder while we incorporate rotations and translations information in the text descriptions. CLIPose achieves state-of-the-art performance on two mainstream benchmark datasets, REAL275 and CAMERA25, and runs in real-time during inference (40FPS).

As the most essential property in a video, motion information is critical to a robust and generalized video representation. To inject motion dynamics, recent works have adopted frame difference as the source of motion information in video contrastive learning, considering the trade-off between quality and cost. However, existing works align motion features at the instance level, which suffers from spatial and temporal weak alignment across modalities. In this paper, we present a Fine-grained Motion Alignment (FIMA) framework, capable of introducing well-aligned and significant motion information. Specifically, we first develop a dense contrastive learning framework in the spatiotemporal domain to generate pixel-level motion supervision. Then, we design a motion decoder and a foreground sampling strategy to eliminate the weak alignments in terms of time and space. Moreover, a frame-level motion contrastive loss is presented to improve the temporal diversity of the motion features. Extensive experiments demonstrate that the representations learned by FIMA possess great motion-awareness capabilities and achieve state-of-the-art or competitive results on downstream tasks across UCF101, HMDB51, and Diving48 datasets. Code is available at https://github.com/ZMHH-H/FIMA.