VGGSounder: Audio-Visual Evaluations for Foundation Models
Researchers have introduced VGGSounder, a new benchmark dataset designed to more accurately evaluate audio-visual foundation models. The existing VGGS dataset has limitations such as incomplete labeling and misaligned modalities, which can distort performance assessments. VGGSounder addresses these issues with comprehensive re-annotations and detailed modality information, allowing for precise analysis of individual modality performance and the impact of combining them. AI
IMPACT Provides a more accurate evaluation tool for audio-visual foundation models, potentially guiding future development.