Interpreting the Linear Structure of Vision-language Model Embedding Spaces

Vision-language models encode images and text in a joint space, minimizing the distance between corresponding image and text pairs. How are language and images organized in this joint space, and how do the models encode meaning and modality? To investigate this, we train and release sparse autoencoders (SAEs) on the embedding spaces of four vision-language models (CLIP, SigLIP, SigLIP2, and AIMv2). SAEs approximate model embeddings as sparse linear combinations of learned directions, or "concepts". We find that, compared to other methods of linear feature learning, SAEs are better at reconstructing the real embeddings, while also able to retain the most sparsity. Retraining SAEs with different seeds or different data diet leads to two findings: the rare, specific concepts captured by the SAEs are liable to change drastically, but we also show that the key commonly-activating concepts extracted by SAEs are remarkably stable across runs. Interestingly, while most concepts are strongly unimodal in activation, we find they are not merely encoding modality per se. Many lie close to - but not entirely within - the subspace defining modality, suggesting that they encode cross-modal semantics despite their unimodal usage. To quantify this bridging behavior, we introduce the Bridge Score, a metric that identifies concept pairs which are both co-activated across aligned image-text inputs and geometrically aligned in the shared space. This reveals that even unimodal concepts can collaborate to support cross-modal integration. We release interactive demos of the SAEs for all models, allowing researchers to explore the organization of the concept spaces. Overall, our findings uncover a sparse linear structure within VLM embedding spaces that is shaped by modality, yet stitched together through latent bridges-offering new insight into how multimodal meaning is constructed.

CAT: A Conditional Adaptation Tailor for Efficient and Effective Instance-Specific Pansharpening on Real-World Data

Pansharpening is a crucial remote sensing technique that fuses low-resolution multispectral (LRMS) images with high-resolution panchromatic (PAN) images to generate high-resolution multispectral (HRMS) imagery. Although deep learning techniques have significantly advanced pansharpening, many existing methods suffer from limited cross-sensor generalization and high computational overhead, restricting their real-time applications. To address these challenges, we propose an efficient framework that quickly adapts to a specific input instance, completing both training and inference in a short time. Our framework splits the input image into multiple patches, selects a subset for unsupervised CAT training, and then performs inference on all patches, stitching them into the final output. The CAT module, integrated between the feature extraction and channel transformation stages of a pre-trained network, tailors the fused features and fixes the parameters for efficient inference, generating improved results. Our approach offers two key advantages: (1) $\textit{Improved Generalization Ability}$: by mitigating cross-sensor degradation, our model--although pre-trained on a specific dataset--achieves superior performance on datasets captured by other sensors; (2) $\textit{Enhanced Computational Efficiency}$: the CAT-enhanced network can swiftly adapt to the test sample using the single LRMS-PAN pair input, without requiring extensive large-scale data retraining. Experiments on the real-world data from WorldView-3 and WorldView-2 datasets demonstrate that our method achieves state-of-the-art performance on cross-sensor real-world data, while achieving both training and inference of $512\times512$ image within $\textit{0.4 seconds}$ and $4000\times4000$ image within $\textit{3 seconds}$ at the fastest setting on a commonly used RTX 3090 GPU.

You Need a Transition Plane: Bridging Continuous Panoramic 3D Reconstruction with Perspective Gaussian Splatting

Recently, reconstructing scenes from a single panoramic image using advanced 3D Gaussian Splatting (3DGS) techniques has attracted growing interest. Panoramic images offer a 360$\times$ 180 field of view (FoV), capturing the entire scene in a single shot. However, panoramic images introduce severe distortion, making it challenging to render 3D Gaussians into 2D distorted equirectangular space directly. Converting equirectangular images to cubemap projections partially alleviates this problem but introduces new challenges, such as projection distortion and discontinuities across cube-face boundaries. To address these limitations, we present a novel framework, named TPGS, to bridge continuous panoramic 3D scene reconstruction with perspective Gaussian splatting. Firstly, we introduce a Transition Plane between adjacent cube faces to enable smoother transitions in splatting directions and mitigate optimization ambiguity in the boundary region. Moreover, an intra-to-inter face optimization strategy is proposed to enhance local details and restore visual consistency across cube-face boundaries. Specifically, we optimize 3D Gaussians within individual cube faces and then fine-tune them in the stitched panoramic space. Additionally, we introduce a spherical sampling technique to eliminate visible stitching seams. Extensive experiments on indoor and outdoor, egocentric, and roaming benchmark datasets demonstrate that our approach outperforms existing state-of-the-art methods. Code and models will be available at https://github.com/zhijieshen-bjtu/TPGS.

AgRowStitch: A High-fidelity Image Stitching Pipeline for Ground-based Agricultural Images

Agricultural imaging often requires individual images to be stitched together into a final mosaic for analysis. However, agricultural images can be particularly challenging to stitch because feature matching across images is difficult due to repeated textures, plants are non-planar, and mosaics built from many images can accumulate errors that cause drift. Although these issues can be mitigated by using georeferenced images or taking images at high altitude, there is no general solution for images taken close to the crop. To address this, we created a user-friendly and open source pipeline for stitching ground-based images of a linear row of crops that does not rely on additional data. First, we use SuperPoint and LightGlue to extract and match features within small batches of images. Then we stitch the images in each batch in series while imposing constraints on the camera movement. After straightening and rescaling each batch mosaic, all batch mosaics are stitched together in series and then straightened into a final mosaic. We tested the pipeline on images collected along 72 m long rows of crops using two different agricultural robots and a camera manually carried over the row. In all three cases, the pipeline produced high-quality mosaics that could be used to georeference real world positions with a mean absolute error of 20 cm. This approach provides accessible leaf-scale stitching to users who need to coarsely georeference positions within a row, but do not have access to accurate positional data or sophisticated imaging systems.

GarmageNet: A Dataset and Scalable Representation for Generic Garment Modeling

High-fidelity garment modeling remains challenging due to the lack of large-scale, high-quality datasets and efficient representations capable of handling non-watertight, multi-layer geometries. In this work, we introduce Garmage, a neural-network-and-CG-friendly garment representation that seamlessly encodes the accurate geometry and sewing pattern of complex multi-layered garments as a structured set of per-panel geometry images. As a dual-2D-3D representation, Garmage achieves an unprecedented integration of 2D image-based algorithms with 3D modeling workflows, enabling high fidelity, non-watertight, multi-layered garment geometries with direct compatibility for industrial-grade simulations.Built upon this representation, we present GarmageNet, a novel generation framework capable of producing detailed multi-layered garments with body-conforming initial geometries and intricate sewing patterns, based on user prompts or existing in-the-wild sewing patterns. Furthermore, we introduce a robust stitching algorithm that recovers per-vertex stitches, ensuring seamless integration into flexible simulation pipelines for downstream editing of sewing patterns, material properties, and dynamic simulations. Finally, we release an industrial-standard, large-scale, high-fidelity garment dataset featuring detailed annotations, vertex-wise correspondences, and a robust pipeline for converting unstructured production sewing patterns into GarmageNet standard structural assets, paving the way for large-scale, industrial-grade garment generation systems.

ChatStitch: Visualizing Through Structures via Surround-View Unsupervised Deep Image Stitching with Collaborative LLM-Agents

Collaborative perception has garnered significant attention for its ability to enhance the perception capabilities of individual vehicles through the exchange of information with surrounding vehicle-agents. However, existing collaborative perception systems are limited by inefficiencies in user interaction and the challenge of multi-camera photorealistic visualization. To address these challenges, this paper introduces ChatStitch, the first collaborative perception system capable of unveiling obscured blind spot information through natural language commands integrated with external digital assets. To adeptly handle complex or abstract commands, ChatStitch employs a multi-agent collaborative framework based on Large Language Models. For achieving the most intuitive perception for humans, ChatStitch proposes SV-UDIS, the first surround-view unsupervised deep image stitching method under the non-global-overlapping condition. We conducted extensive experiments on the UDIS-D, MCOV-SLAM open datasets, and our real-world dataset. Specifically, our SV-UDIS method achieves state-of-the-art performance on the UDIS-D dataset for 3, 4, and 5 image stitching tasks, with PSNR improvements of 9%, 17%, and 21%, and SSIM improvements of 8%, 18%, and 26%, respectively.

Tiling artifacts and trade-offs of feature normalization in the segmentation of large biological images

Segmentation of very large images is a common problem in microscopy, medical imaging or remote sensing. The problem is usually addressed by sliding window inference, which can theoretically lead to seamlessly stitched predictions. However, in practice many of the popular pipelines still suffer from tiling artifacts. We investigate the root cause of these issues and show that they stem from the normalization layers within the neural networks. We propose indicators to detect normalization issues and further explore the trade-offs between artifact-free and high-quality predictions, using three diverse microscopy datasets as examples. Finally, we propose to use BatchRenorm as the most suitable normalization strategy, which effectively removes tiling artifacts and enhances transfer performance, thereby improving the reusability of trained networks for new datasets.

A Bi-channel Aided Stitching of Atomic Force Microscopy Images

Microscopy is an essential tool in scientific research, enabling the visualization of structures at micro- and nanoscale resolutions. However, the field of microscopy often encounters limitations in field-of-view (FOV), restricting the amount of sample that can be imaged in a single capture. To overcome this limitation, image stitching techniques have been developed to seamlessly merge multiple overlapping images into a single, high-resolution composite. The images collected from microscope need to be optimally stitched before accurate physical information can be extracted from post analysis. However, the existing stitching tools either struggle to stitch images together when the microscopy images are feature sparse or cannot address all the transformations of images. To address these issues, we propose a bi-channel aided feature-based image stitching method and demonstrate its use on AFM generated biofilm images. The topographical channel image of AFM data captures the morphological details of the sample, and a stitched topographical image is desired for researchers. We utilize the amplitude channel of AFM data to maximize the matching features and to estimate the position of the original topographical images and show that the proposed bi-channel aided stitching method outperforms the traditional stitching approach. Furthermore, we found that the differentiation of the topographical images along the x-axis provides similar feature information to the amplitude channel image, which generalizes our approach when the amplitude images are not available. Here we demonstrated the application on AFM, but similar approaches could be employed of optical microscopy with brightfield and fluorescence channels. We believe this proposed workflow will benefit the experimentalist to avoid erroneous analysis and discovery due to incorrect stitching.

Whole-Body Image-to-Image Translation for a Virtual Scanner in a Healthcare Digital Twin

Generating positron emission tomography (PET) images from computed tomography (CT) scans via deep learning offers a promising pathway to reduce radiation exposure and costs associated with PET imaging, improving patient care and accessibility to functional imaging. Whole-body image translation presents challenges due to anatomical heterogeneity, often limiting generalized models. We propose a framework that segments whole-body CT images into four regions-head, trunk, arms, and legs-and uses district-specific Generative Adversarial Networks (GANs) for tailored CT-to-PET translation. Synthetic PET images from each region are stitched together to reconstruct the whole-body scan. Comparisons with a baseline non-segmented GAN and experiments with Pix2Pix and CycleGAN architectures tested paired and unpaired scenarios. Quantitative evaluations at district, whole-body, and lesion levels demonstrated significant improvements with our district-specific GANs. Pix2Pix yielded superior metrics, ensuring precise, high-quality image synthesis. By addressing anatomical heterogeneity, this approach achieves state-of-the-art results in whole-body CT-to-PET translation. This methodology supports healthcare Digital Twins by enabling accurate virtual PET scans from CT data, creating virtual imaging representations to monitor, predict, and optimize health outcomes.

Continuous Patch Stitching for Block-wise Image Compression

Most recently, learned image compression methods have outpaced traditional hand-crafted standard codecs. However, their inference typically requires to input the whole image at the cost of heavy computing resources, especially for high-resolution image compression; otherwise, the block artefact can exist when compressed by blocks within existing learned image compression methods. To address this issue, we propose a novel continuous patch stitching (CPS) framework for block-wise image compression that is able to achieve seamlessly patch stitching and mathematically eliminate block artefact, thus capable of significantly reducing the required computing resources when compressing images. More specifically, the proposed CPS framework is achieved by padding-free operations throughout, with a newly established parallel overlapping stitching strategy to provide a general upper bound for ensuring the continuity. Upon this, we further propose functional residual blocks with even-sized kernels to achieve down-sampling and up-sampling, together with bottleneck residual blocks retaining feature size to increase network depth. Experimental results demonstrate that our CPS framework achieves the state-of-the-art performance against existing baselines, whilst requiring less than half of computing resources of existing models. Our code shall be released upon acceptance.