Projects
Explore selected CVI Lab research connecting computational innovation with challenging visual, spatial, and multimodal problems.
A remote-sensing image–text retrieval framework that combines explicit object localization with adaptive coordination between spatial and semantic learning objectives.
A hierarchical mixture-of-experts retrieval framework that promotes both semantic consistency and balanced feature dispersion across remote-sensing image and text representations.
A physics-guided approach for reconstructing gaseous fluid flow, designed to recover more faithful plume structure from challenging visual observations.
AURA improves multi-view contrastive clustering by accounting for sample confidence, cluster proximity, and hubness when constructing and optimizing synthetic universum negatives.
A position paper and benchmark study showing how image generators and reward models can override requests for unconventional, abstract, or deliberately anti-aesthetic imagery.
A remote-sensing retrieval framework that models image–image, image–text, and text–text similarities to identify and reduce false-negative training signals.
A lightweight negative-prompt guidance method that suppresses unwanted concepts by flipping attention value vectors, without retraining the generation model.
A vision–language system that combines video evidence and text prompts to segment gas leaks under supervised and limited-data settings.
A 3D Gaussian head-avatar framework with continuous level-of-detail control, balancing visual quality against rendering cost without retraining.
A progressive matching network that models multiple geographic perspectives and aligns remote-sensing images with textual queries.
A retrieval network that learns compact semantic codes from whole-scene and region-level representations to align remote-sensing imagery with language.
A cross-modal retrieval method that combines spatial-domain visual saliency with frequency-domain cues to emphasize informative remote-sensing content.
A zero-shot gas-leak detection pipeline and synthetic benchmark that combine background subtraction, language-guided object filtering, and promptable segmentation.
A 3D Gaussian head-avatar method designed to accelerate personalization while improving controllability and photorealistic rendering.
A learning-based deformation-transfer method that stylizes new 3D faces from a single artist-created style template without paired training data.
A style-based neural 3D morphable model trained on in-the-wild images for controllable, photorealistic face reconstruction and editing.
A 2D-to-3D face reconstruction method that estimates virtual camera field of view to reduce perspective-driven distortion in recovered facial shape.
A video segmentation framework that combines motion correlations with fine-grained spatial refinement to recover faint gas plumes and their boundaries.
A five-year satellite study of methane concentration patterns over a wastewater treatment plant, including temporal variation and emission-hotspot analysis.
A remote-sensing segmentation transformer that separates scale-dependent features and semantic relationships before integrating them for dense prediction.
An encoder-based reconstruction framework that predicts disentangled HeadNeRF features directly and adds semantic facial-part supervision.
A semantic-segmentation network that decouples high-order contextual relationships to improve dense labeling of complex remote-sensing imagery.
A remote-sensing image–text retrieval network that jointly decouples visual scale information and semantic relationships for more precise matching.
A semantic-segmentation network that represents multi-scale remote-sensing features as interacting graphs to exchange local and global contextual information.
A semantic part-based face model that makes reconstructed 3D faces locally editable while improving their alignment to a source image.
A remote-sensing segmentation framework that decouples multi-scale visual patterns and region relationships before recombining them for dense prediction.
A satellite-image colorization method that uses semantic salience priors to guide plausible colors under strongly imbalanced scene distributions.
A detection pipeline that combines super-resolution with multi-task generative learning to expose small objects in complex, large-scale spatial imagery.
