BROWSE
Detail
Conference Paper
MaterialNet: Multi-scale Texture Hierarchy and Multi-view Surface Reflectance for Material Type Recognition
- Authors
- Dongjin Lee, Hyun-Cheol Kim, Jeongil Seo, Seungkyu Lee
- Issue Date
- 2022-11
- Citation
- British Machine Vision Conference (BMVC) 2022, pp.1-13
- Publisher
- BMVA
- Language
- English
- Type
- Conference Paper
- Abstract
- Material is distinguishing characteristic of real world objects. Recognizing unique texture of certain material type enables improved object detection or semantic segmentation. Incorporating acquired material properties such as surface reflectance of real world objects makes more realistic and richer 3D models in computer graphics. A robot arm essentially requires to recognize the stiffness or roughness of target object for precise and undamaged interaction. Despite the necessities, recognizing material type and its properties from color image is a challenging task. In this work, we propose (1) multi-scale texture hierarchy extraction network (MSTH-Net) encoding view-independent comprehensive multi-scale textures and their hierarchy and (2) multi-view surface reflectance extraction network (MVSR-Net) encoding view-specific features revealing surface reflectance of a material type. Finally, MaterialNet is proposed combining MSTH-Net and MVSR-Net for material type recognition from multi-view color images. Extensive experimental evaluations on six public benchmark datasets show promising performance of proposed method and potential for practical applications.
- KSP Keywords
- 3d model, Benchmark datasets, Color images, Computer graphics, Multi-scale, Multi-view, Object detection, Real-world, Robot Arm, Semantic segmentation, Specific features