Comparative analysis

Select the best architecture from the below

Encoder :

• variations:
  • standard conv2
  • depth-wise separable convolution (1)
  • MobileNet
• initialization:
  • random (1)
  • pre-train as classification on ImageNet
• metric at component level: number of weights, lower is better.

Decoder:

• Fix the "UpConv Block"
• Variations:
  • NNConv3 with interpolation using nearest (1)
  • NNConv3 with pixel shuffle
  • Just pixel shuffle
• Note: Why using convolution during decoding before upsampling.

Loss function:

• train with only the last decoder (decorer_0) (1)
• train with multi-scale depth prediction.

Training method:

• Supervised learning (1)
• Unsupervised learning using stereo images.