Page 46 - 2024S
P. 46
UEC Int’l Mini-Conference No.52 39
highlighting the potential for cross-modal
enhancement in multimedia experiences.
In conclusion, our user study demonstrates
the potential of olfactory-enhanced multimedia
to significantly increase immersion and engage-
ment. The system showed particular strength
in matching strong, distinctive odors like cof-
fee and curry with video content. However, it
also reveals areas for improvement, such as odor
dissipation, intensity calibration, and personal-
ization of olfactory experiences. Future work
Figure 9: Histogram of immersion level scores
for both videos should focus on refining odor presentation tech-
niques and exploring the complex interplay be-
tween different sensory modalities in multimedia
5.4.3 Challenges and Observations contexts.
1. Residual Odors: Some participants noted
lingering scents from previous odor presen- 6 Conclusions
tations, affecting the perception of subse-
quent odors. This highlights the need for Our research presents a novel approach to en-
improved odor dissipation mechanisms in hancing multimedia experiences by integrating
future iterations of the system. odor detection based on semantic analysis of
video subtitles. By leveraging advanced lan-
2. Odor Expectations: Participants often guage models and modern prompt engineering
expected certain smells based on visual cues techniques, we have developed a system capable
(e.g., expecting soy sauce smell during food of predicting relevant odors with remarkable
scenes), even when these weren’t presented. accuracy, achieving over 95% in our quantita-
This underscores the importance of aligning tive evaluations.
olfactory stimuli with strong visual cues for
a coherent experience. The key contributions of our work include:
3. Subtle vs. Strong Odors: While strong, • Advanced Text Analysis: Utilizing
distinct odors like coffee and curry were state-of-the-art large language models with
easily recognized, subtler scents like men- well-crafted prompt engineering to accu-
thol or woods were sometimes missed or rately interpret and predict odors from sub-
misidentified. This suggests a need for care- title contexts.
ful calibration of odor intensities in future • Fine-tuning for Precision: Significantly
studies. improving model performance through
domain-specific training on a specialized
4. Individual Differences: Participants’
nasal conditions and personal odor sensitiv- dataset of 840 text-odor pairs.
ity affected their experiences, emphasizing • Comprehensive Evaluation: Identifying
the importance of accounting for individual the fine-tuned gpt-3.5-turbo as the most
differences in olfactory perception. cost-effective and efficient option through
extensive benchmarking.
5. Cross-modal Interactions: The study
revealed strong interactions between visual, • User-Centric Validation: Demonstrat-
auditory, and olfactory cues. Participants ing increased immersion and engagement
often reported smelling odors that matched in real-world settings using a 13-component
the visual content, even when not present, olfactory display system.
