Large-scale visual understanding, novel-view synthesis and 3D reconstruction
My research involves developing and applying computer vision algorithms that understand, analyze,
and organize large-scale datasets of images and video using deep learning-based techniques.
My recent focus has been on creating computer vision algorithms that can generate and synthesize
3D models of visual scenes from 2D images, using techniques such as multi-view stereo, stable diffusion, and other deep learning models
A Markovian View of Iterative-Feedback Loops in Image Generative Models: Neural Resonance and Model Collapse
2026
Vibhas Kumar Vats, David Crandall, Samuel Goree • ArXiv-preprint
This paper investigates why generative models can degrade when they are repeatedly trained on AI-generated data. We introduce the idea of neural resonance, where iterative feedback pushes models toward a low-dimensional latent structure, and show the conditions under which this happens. We also analyze several model families and describe common patterns of collapse that can help guide future mitigation strategies.
DiffusionGenerative ModelsDeep LearningNeural Resonance
Geometric Constraints in Deep Learning Frameworks: A Survey
2025
Vibhas Kumar Vats, David Crandall • ACM Computing Surveys
This survey examines how geometry can be integrated into deep learning frameworks for depth estimation and related computer vision problems. It presents a taxonomy of common geometry-based constraints, compares existing methods, and discusses future research opportunities.
3D ReconstructionDepth EstimationMulti-View StereoSurvey
HVPUNet: Hybrid-Voxel Point-cloud Upsampling Network
2025
Huyung Ha, Vibhas Kumar Vats, Soon-heung Jung, Alimoor Reza, David Crandall • ICCV
This paper presents an efficient method for point-cloud upsampling from sparse or incomplete 3D data. We propose Hybrid Voxels, which combine the efficiency of voxel-based processing with the precision of point-based representations, and build HVPUNet on top of this design. The resulting framework restores missing geometry, refines surface detail, and achieves strong accuracy with lower computational cost.
3D ReconstructionSuper ResolutionPoint Cloud Upsampling
