dsp-uga / kampf Goto Github PK

As Quinn noted in the project write-up, what we need to do is only the cilia. So I think it's better to check

1. whether the mask is only about cilia. (looking at mask PNG probably won't help because they all seem black.) (The another seems to be no. Because according to Quinn's write-up, there are 3 different labels).
2. if not, then segment out the cilia part and store them as new masks. (should be straightforward.)

Anybody wants to take stabs?

So here are what Quinn mentioned in the project write-up.

As I read it, there are a few things we can do here. The first thing is to figure out what beat frequency exactly is. Also the median filter. Also the optical flow space...

Any ideas?

Better to have a good final version of the notebook

If we have multiple good results, let's find the best of them by taking a vote from them (i.e. blend the results).

Here are my script to do it (preliminary just for 2):

# Blend the results
res_glob = glob('.results/*.png')
res_glob_imgs = array([imread(f) for f in res_glob])
res_glob_2 = glob('.results_2/*.png')
res_glob_imgs_2 = array([imread(f) for f in res_glob_2])
concat_imgs = [res_glob_imgs[i] + res_glob_imgs_2[i] for i in range(len(res_glob_imgs))]

If we want to see the results:

for i in concat_imgs[0]:
    print (i)

Also, we should make sure it's the dtype uint8

need to have this. Any takers?

As I mentioned in the Slack chat, it would be nice to give it a try.

To start with, check the OpenCV document here:
https://docs.opencv.org/3.3.1/d7/d8b/tutorial_py_lucas_kanade.html

Hey Jeremy,

I would like to have a list of the libraries and dependencies to install on the GCC VM. Please also specify versions of each.

Thanks.

What do you all think? Should we calculate the mean and standard deviation of training images so we can normalize the images before training? If so, shall we calculate different mean and std for each video, or calculate mean and std for all dataset?

So as we have read in Quinn Group's paper, to solve the problem of different dimensionalities of input images and masks, they did random crop and then vertical and horizontal flip.

But right now I still can't get this done. What can be easily done is to do a resize of input images and masks (say, put them all to (512, 512)). I'm not sure whether it's good or not

Since there are different sizes of the images, maybe one way to work around that is to change the batch size to one (so it’s guaranteed that PyTorch won’t complain about the inconsistency of the sizes for that batch). Need to test it to find out whether this works! Otherwise I have to the ugly resizing thing.....🙄

Used a tiny image set to start but I found out they are all in different sizes. So I should probably resize them to the same.

Also normalize every image.

Probably get this done no later than tomorrow.