David G. Shatwell — Computer Vision PhD Researcher
David G. Shatwell is a PhD student at the Center for Research in Computer Vision, University of Central Florida. His work centers on geo-localization, 3D vision, and multimodal learning. For any questions, contact him at dshatwell23 [at] gmail (dot) com.
Education and Professional Experience
Selected Publications
TIGeR: A Unified Framework for Time, Images and Geo-location Retrieval
IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) 2026 — David G. Shatwell, Sirnam Swetha, Mubarak Shah
Many real-world applications—digital forensics, urban monitoring, and environmental analysis—demand joint reasoning over visual appearance, location, and time. Beyond conventional geo-localization and capture-time prediction, such tasks increasingly call for retrieving an image taken at the same place as a query but at a chosen target time. The authors formalize this as Geo-Time Aware Image Retrieval and assemble a varied benchmark of 4.5M image–location–time triplets for training plus 85k high-quality triplets for evaluation. They then introduce TIGeR, a multi-modal-transformer model that embeds images, geolocation, and time into one shared geo-temporal space. TIGeR handles flexible single- and multi-modality queries and uses the same representation for geo-localization, time-of-capture prediction, and geo-time-aware retrieval. By preserving location identity under substantial appearance change, it enables retrieval based on where and when a scene is rather than purely visual similarity. Experiments show consistent gains over strong baselines and state-of-the-art methods of up to 16% for time-of-year, 8% for time-of-day prediction, and 14% in geo-time-aware retrieval recall.
VRR-QA: Visual Relational Reasoning in Videos Beyond Explicit Cues
IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) 2026 — Sirnam Swetha, Rohit Gupta, Parth Parag Kulkarni, David G. Shatwell, Jeffrey A. Chan Santiago, Nyle Siddiqui, Joseph Fioresi, Mubarak Shah
Video Question Answering has progressed via multimodal alignment of visual and textual signals, but most existing benchmarks focus on questions answerable from explicit content—actions, objects, and events visible in single frames or short clips. Genuine video understanding requires going beyond what is shown, inferring hidden relationships and contextual cues implied across frames. People do this naturally, integrating partial cues over time to infer motion dynamics, spatial layout, and context, and building a coherent mental model even when relationships are never explicitly depicted. Current benchmarks miss this dimension. To fill the gap, the authors present VRR-QA, a benchmark for Visual Relational Reasoning Beyond Explicit Cues, curated from creative and cinematic videos such as movies that deliberately omit direct depictions of certain events or relations. VRR-QA contains expert-annotated QA pairs drawn from creative clips spanning genres and decades, both live-action and animated. Annotations are deliberately challenging, author-crafted, validated by multiple annotators, and benchmarked against human performance. Extensive evaluations of leading VideoQA models reveal consistent, significant degradation, underscoring reliance on surface-level cues; even the best model falls well short of human baselines at only 64% accuracy. By releasing the dataset and collection framework, VRR-QA offers a rigorous, diverse, reproducible testbed for advancing VideoQA.
Project Page Arxiv Code Dataset
GT-Loc: Unifying When and Where in Images Through a Joint Embedding Space
International Conference on Computer Vision, ICCV 2025 — Oral Presentation (top 0.6%) — David G. Shatwell, Ishan Rajendrakumar Dave, Sirnam Swetha, Mubarak Shah
Timestamp prediction infers when an image was captured from visual information alone, aiding metadata correction, retrieval, and digital forensics. Outdoors, hourly estimates rely on brightness, hue, and shadow positioning, while seasonal change and weather inform date estimation—cues that depend heavily on geographic context and tie timestamp prediction closely to geo-localization. To address this interdependence, the authors propose GT-Loc, a retrieval-based method that jointly predicts capture time (hour and month) and geo-location (GPS coordinates). Separate encoders for images, time, and location align their embeddings in a shared high-dimensional space. Recognizing time's cyclical nature, the method replaces conventional contrastive learning with hard positives and negatives by a temporal metric-learning objective that provides soft targets via pairwise time differences over a cyclical toroidal surface. New benchmarks show the joint optimization surpasses prior time prediction methods, even those given ground-truth geo-location at inference, while achieving competitive results on standard geo-localization and enabling compositional and text-based image retrieval through the unified embedding space.
Project Page Arxiv Poster
Real Time Ore Sorting Using Color and Texture Analysis
International Journal of Mining Science and Technology, 2023 — David Shatwell, Victor Murray, Augusto Barton
Sensor-based ore sorting separates high-grade mineralized rock from low-grade waste to reduce operating costs. While many color-image sorting algorithms exist, few validate results against mineral grades or optimize for real-time classification. This work presents an image-processing and machine-learning algorithm that classifies rocks from a gold and silver mine by grade. It has four stages: image segmentation and partitioning into sub-images; feature extraction using color statistics, principal component analysis, and wavelet texture analysis; sub-image classification with neural networks; and a voting system to determine the rock's overall class. Training used rocks classified by a geologist by mineral content and rocks analyzed in a geochemical laboratory for gold and silver grades. Compared with common classifiers such as convolutional neural networks and support vector machines, the method achieved a Matthews correlation coefficient of 0.961 with processing time under 44 ms, promising for real-time ore sorting.
Research and Engineering Projects
Control Software for Color and 3D Profiling Cameras
Hugo Negreyros, David Shatwell
Unlike area-scan cameras, line-scan cameras capture a single row instead of a full 2D image. This approach brings several advantages over traditional area-scan methods, particularly for image capture.
This site gathers David G. Shatwell's research on geo-localization, 3D vision, and multimodal learning, along with his publications and engineering projects. Use the navigation to browse papers, project pages, code, and his CV.