Material-Based Segmentation of Objects

We present a data-driven proof of concept method for image-based semantic segmentation of objects based on their materials. We target materials with complex radiometric appearances, such as reflective and refractive materials, as their detection is particularly challenging in many modern vision systems. Specifically, we select glass, chrome, plastic, and ceramics as these often appear in real-world settings. A large dataset of synthetic images is generated with the Blender 3D creation suite and the Cycles renderer. We use this data to fine-tune the pre-trained DeepLabv3+ semantic segmentation convolutional neural network. The network performs well on rendered test data and, although trained with rendered images only, the network generalizes so that the four selected materials can be segmented from real photos.