This document discusses FastDepth, a real-time depth estimation model. It begins with prerequisites on dense prediction vs classification networks. The key points section describes depthwise separable convolutions and skip connections. Section 3 describes the FastDepth network architecture, which uses a MobileNet encoder, depthwise separable convolutions, additive skip connections, compiler optimizations, and pruning. Section 4 evaluates these components. The conclusion is that FastDepth achieves real-time performance through its lighter architecture, optimizations, and accuracy. Future work includes updating components like MobileNet and further optimization.
4. Comparison With Classification
Dense Prediction(1/2)
General Classification Network Dense Prediction
• High Resolution -> Low Resolution -> High Resolution
• Encoder + Decoder(>>)
• High Resolution -> Low Resolution -> Class
• Encoder
vs
5. Heavy Decoder on Runtime
Dense Prediction(2/2)
Dense Prediction
Heavy Decoder
• Encoder + Decoder(>>)
29. TODO
[Implementation] MobileNet → MobilNetV2
[Implementation]TVM
[Implementation] NetAdapt
[Study] Why depthwise convolution is slower on cpu
[Study]
31. How It Works
Deconvolution =
Transposed Convolution = Fractionally-Strided Convolution
https://simonjisu.github.io/python/2019/10/27/convtranspose2d.html
https://towardsdatascience.com/types-of-convolutions-in-deep-learning-717013397f4d