Presentation Details
| Scaling Sensory Annotation of Odor Mixtures with a Prior-guided Sensory Annotation Tool Marissa L.Kamarck1, Wesley Qian1, Richard Gerkin1, 2. 1Osmo Labs, PBC, New York, NY, USA.2School of Life Sciences and School of Mathematical and Statistical Sciences, Arizona State University, Tempe, AZ, USA |
Abstract
Data-driven approaches to olfactory prediction require large volumes of high-quality, consistent sensory annotations—particularly for odor mixtures, which dominate real-world olfactory experiences but remain challenging to characterize at scale. At Osmo, we have built Studio, a caption-to-mixture platform for fragrance creation. A critical prerequisite for this effort is the ability to efficiently and reliably annotate complex odor mixtures within a shared, scalable taxonomy of olfactory terms. Traditional sensory annotation of mixtures is limited by low throughput, annotator variability, and cognitive load, especially when mixtures evoke overlapping or ambiguous perceptual qualities. To address these challenges, we developed a novel annotation tool designed to scale taxonomy-based mixture annotation while maintaining data quality and internal consistency. This tool introduces a structured annotation workflow supported by a probabilistic prior derived from a linear mixture model of odor components – using previous annotations or model predictions for each component – building on prior work in mixture perception and representation. This prior provides an initial hypothesis for mixture descriptors, which annotators can then refine based on perceptual experience. We describe the design of this new tool, the process by which priors are generated, and key tradeoffs of the approach including gains in annotation throughput and consistency, potential biases introduced by model-informed priors, and the extent to which human annotators can correct inaccurate priors. Together, these findings highlight a practical path toward collecting large-scale, high-quality mixture annotations to support the next generation of data-driven fragrance creation tools.
No part of this publication may be reproduced, distributed, or transmitted in any form or by any means, including photocopying, recording, or other electronic or mechanical methods, without the prior written permission of the author.
No part of this publication may be reproduced, distributed, or transmitted in any form or by any means, including photocopying, recording, or other electronic or mechanical methods, without the prior written permission of the author.
Content Locked. Log into a registered attendee account to access this presentation.