Rebus puzzles, visual riddles that encode language through imagery, spatial arrangement, and symbolic substitution, pose a unique challenge to current vision-language models (VLMs).Unlike traditional image captioning or question answering tasks, rebus solving requires multi-modal abstraction, symbolic reasoning, and a grasp of cultural, phonetic and linguistic puns.In this paper, we investigate the capacity of contemporary VLMs to interpret and solve rebus puzzles by constructing a hand-generated and annotated benchmark of diverse English-language rebus puzzles, ranging from simple pictographic substitutions to spatially-dependent cues ("head" over "heels").We analyze how different VLMs perform, and our findings reveal that while VLMs exhibit some surprising capabilities in decoding simple visual clues, they struggle significantly with tasks requiring abstract reasoning, lateral thinking, and understanding visual metaphors.

Vision-Language Models (VLMs) excel at visual understanding but often suffer from visual hallucinations, where they generate descriptions of nonexistent objects, actions, or concepts, posing significant risks in safety-critical applications. Existing hallucination mitigation methods typically follow one of two paradigms: generation adjustment, which modifies decoding behavior to align text with visual inputs, and post-hoc verification, where external models assess and correct outputs. While effective, generation adjustment methods often rely on heuristics and lack correction mechanisms, while post-hoc verification is complicated, typically requiring multiple models and tending to reject outputs rather than refine them. In this work, we introduce REVERSE, a unified framework that integrates hallucination-aware training with on-the-fly self-verification. By leveraging a new hallucination-verification dataset containing over 1.3M semi-synthetic samples, along with a novel inference-time retrospective resampling technique, our approach enables VLMs to both detect hallucinations during generation and dynamically revise those hallucinations. Our evaluations show that REVERSE achieves state-of-the-art hallucination reduction, outperforming the best existing methods by up to 12% on CHAIR-MSCOCO and 34% on HaloQuest.

Various studies have been conducted on instance segmentation and made great strides over the past few years. Most recently, instance-specific mask generation via dynamic kernel predictions has shown the significant performance improvement even without bounding boxes as well as anchors. However, this scheme still does not fully consider dynamic properties since the size of the receptive field is not enough to cover the spatially-meaningful range due to memory limitations. Furthermore, the single-fused feature often fails to grasp complicated boundaries for objects of different sizes. In this article, we propose the dynamic residual filtering method with the Laplacian pyramid, which separately restores the global layout and local boundaries of instance masks. Specifically, we firstly apply the Laplacian pyramid-based decomposition scheme to features encoded from the backbone and subsequently restore sub-band mask residuals from coarse to fine pyramid levels. To do this, we design spatially-aware convolution filters to progressively capture the residual form of mask features at each level of the Laplacian pyramid while holding deformable receptive fields with dynamic offset information. This is fairly desirable since global and local properties of mask features can be accurately restored with keeping the spatial flexibility through the invertible process of the Laplacian reconstruction. Experimental results on the COCO dataset demonstrate that our proposed method achieves the state-of-the-art performance, i.e., 42.7% AP. The code and model are publicly available at:https://github.com/tjqansthd/LapMask.

In this letter, we propose an end-to-end stitching network, which takes two images with a narrow field of view (FOV) as inputs, and produces a single image with a wide FOV. Our method estimates multiple homographies to cover the depth differences in the scene and is therefore robust against parallax distortion. In particular, global warping maps are generated using estimated multiple homographies and adjusted by local displacement maps. The final result is made by warping input images multiple times using the warping maps and then merging warped images with the weight maps. Multiple homographies, local displacement maps, and weight maps are generated simultaneously by our stitching network. To train the stitching network, we construct a dataset using the CARLA simulator. Then, using this dataset, our network is trained by end-to-end supervised learning based on appearance matching loss and depth layer loss. In experiments, we show that our method is superior to existing methods both qualitatively and quantitatively. Also, we provide various empirical studies for in-depth analysis as well as the result of the expansion to 360°panoramas.