Noah Stier
I'm a PhD candidate in computer science at UCSB, advised by Tobias Höllerer. My research focuses on 3D scene understanding and reconstruction, with applications in robotics, augmented and virtual reality, and more.
During my PhD I interned with Apple's 3D reconstruction group, and before starting my PhD I worked as a software engineer at Toyon and Procore.
I earned a B.S. from UCLA in computational and systems biology, where I worked with Fabien Scalzo to develop computational methods for medical image analysis.
Email  /  CV  /  Google Scholar
We develop a vision-language 3D reconstruction system, anchoring CLIP features to the reconstructed mesh to support two augmented reality applications on the Magic Leap 2: spatial search with free-text natural language queries, and intelligent object inventory tracking.
We re-structure the core training objective for unsupervised MVS, allowing our networks to learn far better object boundaries for clean and coherent 3D reconstructions without requiring depth/3D ground truth.
We improve the reconstruction of high-frequency content by fixing a fundamental sampling error for volumetric TSDF ground truth. We also leverage depth guidance from MVS, and a new point back-projection architecture, to achieve highly accurate and detailed reconstructions.
We propose neural de-integration to handle SLAM pose updates during online reconstruction from monocular video. To study this problem, we introduce the LivePose dataset of online SLAM pose estimates on ScanNet, the first to include full pose streams with online updates.
We present an open-source software toolkit for identifying, inspecting, and editing tiny objects in multi-megapixel HDR images. These tools offer streamlined workflows for analyzing scientific images across many disciplines, such as astronomy, remote sensing, and biomedicine.
We present a high-quality, annotated DSA vessel segmentation dataset, the largest to-date that is publicly available. We identify semi-automated annotation methods that significantly reduce the labeling cost while preserving the ability to train accurate segmentation models.
We introduce a 3D reconstruction model with a novel multi-view fusion method based on transformers. It models occlusion by predicting projective occupancy, which reduces noise and leads to more detailed and complete reconstructions.
We introduce a deep multi-view stereo network that jointly models all depth maps in scene space, allowing it to learn geometric priors across entire scenes. It iteratively refines its predictions, leading to highly coherent reconstructions.
We present a framework for wide-area detection and recognition in co-incident aerial RGB images and LiDAR, and we source ground-truth semantic labels from OpenStreetMap.
We propose deep neural networks for CT that are equivalent to classical methods at initialization, but can be trained to optimize performance while maintaining convergence properties and interpretability.
We predict the extent of tissue damage following stroke, using the cerebral hypoperfusion feature observed in MRI after stroke onset. We show improved performance using a CNN vs. the standard single-voxel regression model.
We train a CNN to classify regions of hyperperfusion based on cerebral blood flow (CBF) observed in MRI after stroke onset, approaching the accuracy of human experts.
Website design from Jon Barron.