Towards Improving Monocular Depth Estimation for Mobile Devices

We propose a new method for fast and accurate monocular depth estimation for embedded systems. Our objective was to improve the FastDepth model, which achieves state-of-the-art per-formance for mobile applications. To do this, different loss functions were tested for training the model and more advanced light-weight CNNs were used as encoders. Even though switching from the default L2 to the BerHu loss function had as a result only a minor improvement in the quality of the generated depth maps, we managed to enhance significantly the performance of Fast-Depth, by using an EfficientNet based CNN as the encoder. These changes had but a minuscule effect on the model’s size and latency, which is why we argue that our approach is an improvement for mobile applications

FastDepth Depth-maps with our Encoder:

Figure: 1st column: RGB input image, 2nd column: ground truth depth map, 3rd column: generated depth map

	Size	No. Params	Avg. Latency	RMSE (mm)	δ < 1.25
Original FastDepth	31MB	4M	9msecs	796	0.603
EfficientDepth (ours)	40MB	6.5M	23msecs	678	0.688

Using our model to generate 3D images with 3d photo inpainting