This work presents a computer vision framework for analyzing the ripening process of cranberry crops using both aerial drone and ground-based imaging across a full growing season. By leveraging vision transformers (ViT) and UMAP dimensionality reduction, the framework enables interpretable visualizations of berry appearance and quantifies ripening trajectories. The approach supports precision agriculture tasks such as high-throughput phenotyping and crop variety comparison. This is the first visual framework for cranberry ripening assessment, with potential impact across other crops like wine grapes and olives.

Vision-on-the-bog is framework for smart agriculture that enables real-time decision-making by monitoring cranberry crops. It performs instance segmentation to count sun-exposed cranberries at risk of overheating and predicts internal berry temperature using drone and sky imaging. A weakly supervised segmentation method reduces annotation effort, while a differentiable model jointly estimates solar irradiance and berry temperature. These tools support short-term risk assessment to inform irrigation decisions. The approach is validated over two growing seasons and can be extended to crops such as grapes, olives, and grain.

This work introduces the Landmark-Aware Visual Navigation (LAVN) dataset to support supervised learning of human-centric exploration and map-building policies. The dataset includes RGB-D observations, human point-clicks for navigation waypoints, and annotated visual landmarks from both virtual and real-world environments. These annotations enable direct supervision for learning efficient exploration strategies and landmark-based mapping. LAVN spans diverse scenes and is publicly released with comprehensive documentation to facilitate research in visual navigation.

This paper presents DEVI, a self-supervised, class-agnostic object detector trained on egocentric video without annotations or pretraining. It uses multi-view and scale-regression losses—motivated by appearance-based cues—to learn dense, category-specific features. The learned cluster residual module enables robust object localization in complex environments. DEVI achieves significant gains in detection metrics on the Ego4D and EgoObjects datasets while using a lightweight, end-to-end architecture.

This paper presents a single-stage method for weakly‑supervised semantic segmentation using point annotations to generate reliable pseudo‑masks on the fly, eliminating the need for pre‑training, multi‑stage pipelines, or refined CAMs. It achieves state‑of‑the‑art performance across challenging real‑world benchmarks such as CRAID, CityPersons, ADE20K and CityScapes, showing advantages in both accuracy and generalizability. The method is trained from scratch, handles complex scenes, and outperforms multi‑stage methods that rely on pre‑trained networks. Its innovation lies in combining expanding distance fields with pixel‑adaptive convolutions to refine masks at each training step.

We present a system that segments buildings and bridges from terrain using multiview satellite imagery, and reconstructs low‑polygon textured 3D meshes with preserved sharp edges. Our method focuses on semantic segmentation and regularized surface extraction, evaluated on the IARPA CORE3D dataset. A web‑based AWS application enables users to deploy algorithms and visualize results, with both code and interface released as open-source. This addresses challenges of dense, “melted” point‑cloud models and a lack of semantic separation.

H2O‑Net presents a self‑supervised deep-learning method for flood segmentation by bridging domain gaps between low‑resolution government satellite imagery and high‑resolution commercial satellite/drone imagery using adversarial adaptation and label refinement. It synthesizes SWIR-like signals to enhance water detection, then refines coarse masks via a self‑supervised mechanism, requiring no manual annotations. The model achieves ~10–12% improvements in pixel accuracy and mIoU over state-of-the-art methods and generalizes well across sensors. The approach supports rapid flood mapping and demonstrates real-world viability for disaster response.

This work presents a novel pipeline for material segmentation using multi-view satellite imagery, addressing the challenge of combining low-resolution panchromatic and multispectral data. A specialized CNN architecture is introduced that learns both spatial and angular reflectance cues for improved material classification. The model is trained and evaluated using the SpaceNet 6 dataset and demonstrates significant gains over baseline methods in segmenting roads, buildings, and natural terrain. The study underscores the importance of angular diversity in satellite viewpoints for material understanding.

This work introduces a multi-view approach to terrain material classification by analyzing appearance variations from different viewpoints. A convolutional neural network is trained to learn discriminative features across angular changes, improving recognition accuracy on outdoor surfaces. Benchmark results on a novel dataset show significant performance gains over single-view baselines. The method highlights the importance of viewpoint diversity in robust material perception.

This paper presents a method for photo-realistic facial texture transfer that maintains high-frequency identity details while adapting global appearance. The approach uses a hybrid deep learning model to align features between a source and target domain. Applications include facial editing, animation, and virtual reality content synthesis.

Deep TEN introduces an end-to-end trainable texture encoding layer within CNNs to enhance texture recognition. The model aggregates deep local features into a robust global descriptor, significantly improving texture classification performance. It sets new benchmarks on multiple texture and material datasets.

We introduce a novel method for using reflectance to identify materials. Reflectance offers a unique signature of the material but is challenging to measure and use for recognizing materials due to its high-dimensionality. In this work, one-shot reflectance of a material surface which we refer to as a reflectance disk is capturing using a unique optical camera. The pixel coordinates of these reflectance disks correspond to the surface viewing angles. The reflectance has class-specific stucture and angular gradients computed in this reflectance space reveal the material class. These reflectance disks encode discriminative information for efficient and accurate material recognition. We introduce a framework called reflectance hashing that models the reflectance disks with dictionary learning and binary hashing. We demonstrate the effectiveness of reflectance hashing for material recognition with a number of real-world materials.

This work introduces a novel method to capture fine-scale surface geometry, or relief texture, which is critical for realistic rendering and accurate object modeling. Traditional scanners often miss such detail—especially on specular or partially specular surfaces—due to light scattering. We develop a custom imaging device using a concave parabolic mirror that captures multiple viewing angles in a single image, enabling high-resolution recovery of both surface shape and spatially varying reflectance. Unlike laser scanning, this approach is tailored to specular materials and captures both texture shape and color simultaneously.

In this work, we investigate the visual appearance of real-world surfaces and the dependence of appearance on the geometry of imaging conditions. We discuss a new texture representation called the BTF (bidirectional texture function) which captures the variation in texture with illumination and viewing direction. We present a BTF and BRDF database with image textures from over 60 different samples, each observed with over 200 different combinations of viewing and illumination directions. Both of these unique databases are publicly available and have important implications for computer graphics.

Detection of cracks on bridge decks is a vital task for maintaining the structural health and reliability of concrete bridges. Robotic imaging can be used to obtain bridge surface image sets for automated on-site analysis. We present a novel automated crack detection algorithm, the STRUM (spatially tuned robust multifeature) classifier, and demonstrate results on real bridge data using a state-of-the-art robotic bridge scanning system. A crack density map for the bridge mosaic provides a computational description as well as a global view of the spatial patterns of bridge deck cracking. The bridges surveyed for data collection and testing include Long-Term Bridge Performance program's (LTBP) pilot project bridges at Haymarket, VA, USA, and Sacramento, CA, USA.

Ground penetrating radar (GPR) is used to evaluate deterioration of reinforced concrete bridge decks based on measuring signal attenuation from embedded rebar. The existing methods for obtaining deterioration maps from GPR data often require manual interaction and offsite processing. In this paper, a novel algorithm is presented for automated rebar detection and analysis. We test the process with comprehensive measurements obtained using a novel state-of-the-art robotic bridge inspection system equipped with GPR sensors. The algorithm achieves robust performance by integrating machine learning classification using image-based gradient features and robust curve fitting of the rebar hyperbolic signature. High rates of accuracy are reported on real data that includes thousands of individual hyperbolic rebar signatures from three real bridge decks.

This work introduces Eigenbiome, a statistical model that links photographic skin appearance to underlying microbial community features using unsupervised dimensionality reduction. By capturing texture and color patterns in a low-dimensional “appearance space,” the framework infers likely microbiome compositions without direct sampling. Experiments on real facial and forearm imagery show strong correlations between appearance-based vectors and measured bacterial profiles. The approach opens a new avenue for non-invasive skin health monitoring and microbiome estimation.

This paper presents a registration framework for aligning multimodal (reflectance and fluorescence) and time‑lapse skin images at subpixel resolution despite motion variations. The system automatically detects micro‑features like pores and tracks them over weeks to monitor lesion evolution. Results show high alignment accuracy, improving downstream quantitative skin analysis. This robust alignment method enhances data quality for dermatological research and clinical monitoring.

This work introduces a deep model for learning skin micro- and macro-texture features, combining local descriptors and global mapping for clinical classification of dermatological conditions. The system incorporates a texture-encoding branch to capture fine geometric skin patterns, validated on clinical and consumer images. It achieves high accuracy in classifying conditions like eczema and acne, outperforming baseline texture descriptors. This model supports interpretability by highlighting clinically relevant texture regions.