Daily Arxiv

This is a page that curates AI-related papers published worldwide.
All content here is summarized using Google Gemini and operated on a non-profit basis.
Copyright for each paper belongs to the authors and their institutions; please make sure to credit the source when sharing.

OE3DIS: Open-Ended 3D Point Cloud Instance Segmentation

Created by
  • Haebom

Author

Phuc DA Nguyen, Minh Luu, Anh Tran, Cuong Pham, Khoi Nguyen

Outline

This paper presents the Open-Kind 3D Instance Segmentation (OE-3DIS) problem, which enables novel object segmentation without predefined class names. Existing open-vocabulary 3D instance segmentation (OV-3DIS) methods suffer from the limitation of relying on predefined class names during testing; OE-3DIS alleviates this limitation. We build a robust baseline model by leveraging the OV-3DIS approach and a 2D multimodal large-scale language model, and evaluate its performance using a novel Open-Kind Score and a standardized AP score, which assess the semantic and geometric quality of predicted masks and their associated class names. On the ScanNet200 and ScanNet++ datasets, we achieve significant performance improvements over the baseline model, and even outperform the previous state-of-the-art OV-3DIS method, Open3DIS.

Takeaways, Limitations

•
Takeaways:
◦
By defining the OE-3DIS problem, which enables 3D instance segmentation without predefined class names, and presenting a powerful baseline model and evaluation metrics for it, we contribute to the development of more autonomous 3D object recognition systems.
◦
Improving 3D instance segmentation performance using 2D multimodal large-scale language models.
◦
A new open kind score enables comprehensive assessment of semantic and geometric quality.
◦
Achieve performance that surpasses existing top-performing models.
•
Limitations:
◦
Further experiments are needed to evaluate the generalization performance of the proposed method.
◦
Additional performance evaluations on various 3D datasets are needed.
◦
Due to the high reliance on 2D multimodal large-scale language models, the limitations of the models may affect the performance of OE-3DIS.
👍
Made with Slashpage