This paper aims at generating anomalous images and their segmentation labels to address the lack of real-world anomaly samples and privacy issues. Departing from conventional approaches that use masks solely to guide the generation of anomaly images, we propose a dual-branch training strategy for the generative model. This strategy enables the simultaneous production of anomaly images and masks, with an alignment regularization loss that ensures the coherence between the generated images and their masks. During inference, only the image-generation branch is activated to produce synthetic samples for training the downstream segmentation model. Furthermore, we propose to integrate the well-trained generative model into the training of segmentation models, utilizing a generative feedback loss to refine the segmentation model's performance. Experiments show our method's IoU metrics exceed previous methods by 5.03%, 5.68% and 16.63% on Real-IAD (industrial), polyp (medical), and Floor Dirty (indoor) datasets. The code is publicly accessible at https://github.com/huan-yin/anomaly-alignment.
1 February 2026 in Tsinghua Science and Technology
Open-world object detection (OWOD) is a challenging task requiring models to detect both known and unknown objects while incrementally learning from new data. Current OWOD methods typically label regions with high objectness scores as unknown objects, relying heavily on known object supervision, leading to label bias. To address this, we propose object reconstruction error modeling, using object-level semantic information for unsupervised foreground and background modeling. Additionally, we introduce an unsupervised proposal generation method, leveraging segment anything model’s zero-shot learning to generate pseudo-labels for unknown objects. However, classifiers trained on known categories tend to bias toward them during inference. To resolve this, we propose a location-enhanced network, reframing classification as a location quality prediction task. Our method achieves a significant 37% improvement in unknown category recall (52.1%) on the Microsoft common objects in context (MS-COCO) dataset, outperforming previous state-of-the-art methods while maintaining competitive performance on known objects. Furthermore, it surpasses deformable detection transformer (DETR)-based models, achieving 10.95 frames per second, with a speed advantage over faster region-based convolutional neural network (Faster R-CNN)-based methods.