Efficient Interactive Annotation of Segmentation Datasets

Reimplementation of Efficient Interactive Annotation of Segmentation Datasets with Polygon-RNN++.
This repository contains two main programs:

• Reproduction of the entire paper
  I used the official model that has been trained,and some code is based on official pytorch reimplementation.
• CNN image feature extractor
  This part is done by me.

Reproduction of the entire paper

dependencies:

• python==3.5.4
• tensorflow==1.3.0
• scikit-image==0.14.2
• numpy==1.14.2
• matplotlib==2.2.2
• tqdm==4.19.9

usage:

1. Dowload the trained model from www.cs.toronto.edu/polyrnn/models/$FILENAME .
2. Unzip it and put the subfolder in the empty mates，at this step you may change their name to match paths in code.
3. Run the rnn_main.py,that will take some time.
4. View the output JSON files and tagged images in the output folder.

output
Test case:

Reduced subimages to match 224*224 rgb:

Testing effect(use RNN or RNN+GGNN):

Overall effect:

Performance:

Feature Extractor

This part of the program mainly uses Keras.

Main Idea:

• Build a CNN model without pooling or FC layers.

We also remove the original average pooling and FC layers

• Construct a ResNet-50 layer model based on reference[13] and reference[7]

we follow [7] and modify the ResNet-50 architecture [13] by reducing the stride of the network and introducing dilation factors.

DeepLab

Batch Normalization

Additional Dependencies:

• six==1.12.0
• Keras==2.2.2

Usage:
Just run extractor_main.py, the result will be presented to you in a Dialog.

Output
test case:

effect: