HOLA: Holistic Multi-Modal Alignment for Open-Set 3D Recognition
Researchers have developed HOLA, a new method for open-set 3D recognition that improves generalization to unseen categories. HOLA aligns 3D point clouds with multiple images and textual descriptions to achieve a more comprehensive understanding of objects. The approach utilizes a novel decoupled multi-positive contrastive loss function that focuses on challenging negatives and avoids issues with multiple positives. Additionally, a lightweight text adapter is employed to bridge the domain gap between web captions and curated annotations, enabling effective use of large-scale unsupervised text data. AI
IMPACT Improves generalization for 3D recognition models, potentially enabling more robust AI systems in real-world scenarios with novel objects.