Boston, Massachusetts, United States
• Initiated the first iteration of material semantic segmentation and achieved 88% mean pixel accuracy with limited synthetic data
• Developed material semantic segmentation network with U-Net, Fully Convolutional DenseNet, and DeepLabV3+ architecture
• Improved inference pipeline by adopting DeepLabV3+ network which runs inference 3.5x faster than DenseNet network
• Collaborated cross-functionally with visual artists and 3D operations team to decide the project scope and data synthesis pipeline
• Conducted data analysis with visualization, designed data synthesis algorithm, designed human-in-the-loop evaluation methods
