V$^{2}$-SAM: Marrying SAM2 with Multi-Prompt Experts for Cross-View Object Correspondence

The introduction of DinoLizer, a model based on DINOv2, aims to enhance the localization of manipulated regions in generative inpainting. By utilizing a pretrained DINOv2 model on the B-Free dataset, it incorporates a linear classification head to predict manipulations at a granular patch resolution, employing a sliding-window strategy for larger images. This method shows superior performance compared to existing local manipulation detectors across various datasets.

A novel approach to Video Frame Interpolation (VFI) has been introduced, focusing on enhancing motion estimation accuracy by utilizing Region-Distinguishable Priors (RDPs) derived from the Segment Anything Model 2 (SAM2). This method aims to address the challenges of ambiguity in identifying corresponding areas in adjacent frames, which is crucial for effective interpolation.

A new method called DWFF-Net has been developed to identify multi-scale farmland system habitats using an adaptive dynamic weight strategy. This approach addresses the shortcomings of existing habitat classification systems by providing a comprehensive dataset of ultra-high-resolution remote sensing images that categorize cultivated land into 15 distinct habitat types.

A new model called Vision-Language Enhanced Semi-supervised Segmentation Assistant (VESSA) has been introduced to improve semi-supervised medical image segmentation by integrating vision-language models (VLMs) into the segmentation process. This model aims to reduce the dependency on extensive expert annotations by utilizing a two-stage training approach that enhances visual-semantic understanding.

The Segment Anything Model 2 (SAM2) has undergone systematic evaluation for its application in surgical video segmentation, revealing its potential for zero-shot segmentation across various surgical procedures. The study assessed SAM2's performance on nine surgical datasets, highlighting its adaptability to challenges such as tissue deformation and instrument variability.

A new study introduces the Interleaved Multi-Domain Identity Curriculum (IMIC), enabling models to perform object recognition, face recognition from varying image qualities, and person recognition in a unified embedding space without significant catastrophic forgetting. This approach was tested on foundation models DINOv3, CLIP, and EVA-02, demonstrating comparable performance to domain experts across all tasks.

Recent advancements in artificial intelligence have led to the development of NeuroVFM, a generalist neuroimaging model trained on 5.24 million clinical MRI and CT volumes. This model was created through a novel approach called health system learning, which utilizes uncurated data from routine clinical care, addressing the limitations faced by existing AI models that lack access to private clinical data.

A new approach to damage assessment in conflict zones has been introduced through the CSD framework, which utilizes a pre-trained DINOv3 model and a multi-scale cross-attention difference siamese network (MC-DiSNet). This method addresses challenges such as high intra-class similarity and ambiguous semantic changes in damaged areas, which often share similar architectural styles and exhibit blurred boundaries.