MUMBAI, India, Sept. 22 -- Intellectual Property India has published a patent application (202621092909 A) filed by Indian Institute Of Technology Bombay on July 31, 2026, for System Apparatus And Method For Hyperbolic Source-Only Domain-Generalized Category Discovery.
Inventors include Gupta, Divyam; Banerjee, Biplab; and Rathore, Vaibhav.
The application for the patent was published on September 18, 2026, under issue no. 38/2026.
Abstract: An image-processing apparatus comprises one or more hardware processors and memory storing instructions that configure the apparatus to train a visual recognition model using labeled source-domain images without target-domain images during training. The apparatus generates source-derived synthetic domains, selects one or more synthetic domains based on a domain-diversity score, extracts visual features using a visual encoder, and maps the visual features into a Poincaré-ball embedding space to generate hyperbolic feature embeddings. Boundary-positioned known-category prototypes, Tangent CutMix pseudo-novel embeddings, penalized Busemann alignment, hybrid hyperbolic contrastive loss, adaptive outlier repulsion, and learnable curvature are used to configure the visual recognition model. During inference, target-domain images are mapped into the Poincaré-ball embedding space and categorized into known visual categories or grouped into one or more novel visual categories.
Disclaimer: Curated by HT Syndication.
