Skin segmentation performs bad on boudaries

[chrigui94]

Hi all, I am using Mask RCNN for skin segmentation in medical images. may dataset is composed of:

• training: 700 images + json annotations
• validation : 150 images + json annotations My config file is..

GPU_COUNT = 1
    IMAGES_PER_GPU = 2
    NUM_CLASSES = 1 + 1 #one for skin and the other for background  
    #TRAIN_ROIS_PER_IMAGE = 64 # default=200
    STEPS_PER_EPOCH = 100
    #VALIDATION_STEPS = 30
    DETECTION_MIN_CONFIDENCE = 0.8
    #USE_MINI_MASK = True
    EPOCHS=100
    LEARNING_RATE = 0.0001
    layers='all'

the problem is that the mask is not completed on the boundaries as shown on this image: Is there any proposition how can I improve this segmentation?

[codeyogi911]

You might want to try: USE_MINI_MASK = False Also the output mask is just 28x28 in the default config file You might want to change this as well for more accurate mask

[codeyogi911]

Also check out #635