milani / science-bowl-2018

My solution to science bowl 2018

Hi there 👋

science-bowl-2018's Issues

Some suggestions to improve your score

Hi, thanks for your code. I checked your current score about 0.36 LB. It is too low than standard implementation of mask rcnn ( I got 0.41). I would like to suggest you some schemes:

1. Randomly crop training image into some fixed size such as 128x128 then train on the size. Don't add the zero padding to a larger size such as 1024x1024, it may lost the image distribution. In the inference, you can split the testing image into sub-images of 128x128 and combine together to make final result (I did not complete the combination step)
2. Set larger RPN_TRAIN_ANCHORS_PER_IMAGE such as 512
3. Use smaller network such as resnet-50 instead of resnet-101
4. Augmentation data as you did

This is my setting to achieve above score

IMAGES_PER_GPU = 2 
GPU_COUNT = 1
NUM_CLASSES = 1 + 1 # Background + Nuclei 
USE_MINI_MASK=True 
MINI_MASK_SHAPE = (56,56) 
RPN_ANCHOR_SCALES = (8, 16, 32, 64,128) 
RPN_ANCHOR_RATIOS = [0.25, 0.5, 1, 2, 4] 
RPN_NMS_THRESHOLD = 0.7 
IMAGE_CROP_WIDTH=128
IMAGE_CROP_HEIGHT=128
DETECTION_MAX_INSTANCES = 200 
RPN_TRAIN_ANCHORS_PER_IMAGE = 512
RPN_NMS_THRESHOLD = 0.6 
MAX_GT_INSTANCES = 200 
DETECTION_MIN_CONFIDENCE = 0.6