• Improve depth estimation model by constructing multi-tasking framework with semantic segmentation task, which brings higher accuracy and more clear boundary of objects in visual quality.

• Quantize depth estimation model without losing much performance and make the quantized model deployable on Qualcomm Snapdragon in real-time.

• Build up monocular depth estimation framework including multi-dataset(e.g. synthetic and real-world dataset) and different training settings (e.g. monocular/stereo based, self-supervision) to incorporate the existing pipeline.

• Implement a face frontalization algorithm to improve face recognition efficiency.