Week #11

07 Jan 2019

Experiments and coding

• After the observation of the overfitting in the network after training, I decide to increase the weight decay.

Number	Data source	Training Sample	epoch of training	Batch size	learning rate	cropped size	momentum	Weight decay	Mimimum loss attained	the result
01	KITTI	160	20	8	0.05, every 5 epoch divided by 5	256	0.9	0.0005	0.608	good
02	KITTI	160	100	8	0.05, every 5 epoch divided by 5	256	0.9	0.0005	0.612	good
03	KITTI	160	20	12	0.05， every 5 epoch divided by 5	256	0.9	0.0005	0.617	good
04	KITTI	480	10	1	0.05, every 5 epoch divided by 5	300	0.9	0.0005	0.604	good
05	KITTI	160	10	1	0.05, every 5 epoch divided by 5	original size	0.9	0.0005	0.595	good
06	KITTI	160	10	2	0.05, every 5 epoch divided by 5	256，1024	0.9	0.0005	0.601	good
07	KITTI	160	10	2	0.05, every 5 epoch divided by 5	256,1024	0.9	0.005	0.606	good
08	KITTI	160	10	1	0.05, every 5 epoch divided by 5	original size	0.9	0.05	0.685	bad
09	KITTI	160	10	1	0.05, every 5 epoch divided by 5	original size	0.9	0.005	0.614	good

• Add the pixel mapping for the mask, convert the [0,255] pixel value to classes. For KITTI dataset, found 29 classes in total. And for BMCC dataset, there are 3 classes.

Paper review

Deep Neural Netwoks Segment Neuronal Membrances in Electron Microscopy Images

Contribution:

• Use a window center one pixel to predict the class of this pixel, i.e. a patch based image Segmentation
• Use Foveation and nonuniform sampling to disregard the fine details and the margin part of the window
• Use several networks with different architecture and average the output of these networks
• Win the ISBI challenge 2012

Drawbacks:

• It’s slow because for every patch the network must run separately
• There is a trade-off between the localization information and the context information, i.e. if the patch size is bigger, the context information is larger while need more max-pooling layers, which reduce the localization information.

My understand of the image segmentation

Image classification

Goal: Given an image, assign a class label to the image.

Image segmentation

Goal: Give an image, assign a mask to the image. The mask represents the score value for every pixel in the image. i.e. The image segmentation problem can be considered as the pixel classification promblem.

MRI images received

Received 330 cases of MRI images. The next step should be playing with the real project.