Monocular depth estimation on embedded devices

Abstract

In this work, a lightweight monocular depth estimation network capable of running on hardware with limited resources was developed. The network produces dense depth maps in indoor environments from low resolution images captured by a single camera. The network utilizes an encoder-decoder based architecture with skip connections. A modified MobileNetV2 was used as the encoder, and the decoder comprised upsampling layers, skip-connection layers and convolutional layers. The network was trained for indoor use cases using NYU Depth V2 dataset and for outdoor use cases using KITTI dataset using a loss function which utilizes structural similarity. The model's parameters were fully quantized to 8 bits from the original 32 bits to further reduce memory requirements. The network was tested using OPENMV H7 microcontroller board and was successfully able to produce depth maps from images captured by the board's camera in real-time.