This part contains the details of background-removal and 5-body-part segmentation processes of moth specimen images from TESRI(Taiwan Endemic Species Research Institute).

• Source data: BRCAS with labels.
• Target data: TESRI without labels.

### pipenv
$ cd /to/working/directory/
$ pipenv sync
$ pipenv shell
  
### get model
$ bash get_models.sh


### train - unsupvised
$ python3 Unsup_train.py --XX_DIR=/path/to/image --SAVEDIR=/path/to/save/ --minLabels=/minimum/number/of/labels/ --gpu='gpu_id'

### train - supervised
$ python3 Sup_train.py --XX_DIR=/path/to/image/ --YY_DIR=/path/to/groundtruth/ --SAVEDIR=/path/to/save/ --num_class= number of output class --gpu='gpu_id'

### predict - background removal
$ python3 Sup_predict_rmbg.py --XX_DIR=/path/to/image --model_dir=/path/to/checkpoint/ --gpu='gpu_id'

### predict - 5-body-part segmentation
$ python3 Sup_predict_5comps.py --XX_DIR=/path/to/image --model_dir=/path/to/checkpoint/ --gpu='gpu_id'

• Elevation_prediction.ipynb builds the ResNet model to predict species elevation.
• Feature_Extraction_and_Assemblage_Clustering.ipynb extracts deep features and then clusters moth species into several assemblages.
• Postprocess.ipynb postprocesses background-removal model results(find_contour and condition random field).
• 5comps_output_process.ipynb processes 5-comps model results to generate final images.
• Visualize.ipynb visulizes some samples of different background-removal steps and 5-comps results.

Here we show some sample results of the complete processes, you could find more samples in result_sample The first row of each species are what we've tried to remove background. Then we chose the best one for each moth to segment its 5-body parts. The second row are the 5-body-part result of that moth.

This repository reuses code from pytorch-unsupervised-segmentation by kanezaki and Data_Science_Bowl_2018 by RaoulMa. Many thanks to all the contributions!