More semantic segmentation models support

[ibelem]

• List models
• Justification (in internal)

[BruceDai]

• FCN8 (OpenVino)

1. Paper: Fully Convolutional Networks for Semantic Segmentation
2. Could convert from fcn8s-heavy-pascal.caffemodel by OpenVino Model Optimizer, converted OpenVino model file size: 537.9MB
3. This model is used by Image Segmentation C++ Demo of OpenVino ToolKit
4. Need implement Deconvolution, Crop ops to support this model

To implement Deconvolution and Crop ops by WebGL of tensorflow, Deconvolution could map to tf.layers.TRANSPOSE_CONV_2D, and Crop map to tf.layers.CROPPING2D.

• DUC (ONNX)

1. Paper: Understanding Convolution for Semantic Segmentation
2. Model file: ResNet101_DUC_HDC size: 260.7MB
3. mIOU: 81.92%
4. OnnxModelImporter could handle the ops of this model, while there's some problem on investigating and fixing.

DUC is based on cityscapes, it isn't OK for Video conference Scene, skip this model.

• Mask R-CNN This model was mentioned on 2.1.2. Semantic Segmentation Use Case of WebML CG It's for object instance segmentation that detects 80 different classes. Paper: Mask R-CNN

[BruceDai]

• Mask R-CNN2GO Reference Enabling full body AR with Mask R-CNN2Go