Conference proceeding
Cosmos: Coherent Scene with Multiple Objects Reconstruction
2024 IEEE International Conference on Acoustics, Speech, and Signal Processing Workshops (ICASSPW), pp 660-664
14 Apr 2024
Abstract
We present the Coherent Scene with Multiple Objects (CoSMOs) framework for high-quality 3D mesh reconstruction of indoor environments from a single image. Recent advances in image-based 3D reconstructions have shown promising results for object-level reconstruction tasks. However, existing methods for single-image 3D scene reconstruction still fail to accurately model large-scale environments with many objects due to the loss of depth information in 2D image and high feature ambiguity around unseen areas. In this work, we leverage vision transformers to better capture depth-wise 3D context and use recently developed diffusion models to further refine blurry results from such ambiguity. Extensive quantitative and qualitative analyses demonstrated the effectiveness of our approach which achieves state-of-the-art performances, surpassing existing methods in terms of quality and fidelity.
Metrics
15 Record Views
Details
- Title
- Cosmos: Coherent Scene with Multiple Objects Reconstruction
- Creators
- Byoungsung Lim - Korea UniversityDavid Han - Drexel University
- Publication Details
- 2024 IEEE International Conference on Acoustics, Speech, and Signal Processing Workshops (ICASSPW), pp 660-664
- Publisher
- IEEE
- Grant note
- Korea Institute for Advancement of Technology (10.13039/501100003661)
- Resource Type
- Conference proceeding
- Language
- English
- Academic Unit
- Electrical and Computer Engineering
- Scopus ID
- 2-s2.0-85202431209
- Other Identifier
- 991021930833604721