New method uses VLM and diffusion to create labeled training data from unlabeled photos

Researchers have developed a novel method for generating labeled training data for object detectors using a small set of unlabeled photographs. This approach leverages vision-language models to create 3D vegetation scenes from single images, allowing for the automatic generation of bounding boxes and segmentation masks. A diffusion model is then used to re-texture these scenes, with a 'mask-lock' parameter controlling how much the diffusion process affects the target objects. This synthesized data, when used to train a standard detector, performed comparably to detectors trained on larger, real-world datasets from different sites, demonstrating effective unsupervised site adaptation for tasks like humanitarian demining. AI

IMPACT Enables creation of labeled training data for object detection with minimal human annotation, potentially accelerating development in specialized domains.

RANK_REASON The cluster contains a research paper detailing a new methodology for data generation in computer vision. [lever_c_demoted from research: ic=1 ai=1.0]

Read on arXiv cs.CV →

• 3D object meshes
• arXiv
• detector
• Humanitarian Demining Center
• vision-language model

• paper
• product

AI-generated summary · Google Gemini · from 1 sources. How we write summaries →