normanrz / face-vid Goto Github PK

Currently we are replacing every occurrence of 0 in the image with the mean, converting every black pixel to a mean pixel. This obviously also changes the color of pixels that are not part of the mask.

Possible solution:

• get rid of resizing (probably not a good idea anyway). Already crop a properly sized image from the face video.
• draw the ellipse with image independent parameters (we are actually doing that already, but the problem is the resizing of the frames after painting the ellipse). First in black and in the second pass with the mean color

According to BVLC/caffe#1470 HDF5 files must not be larger than 2GB unzipped.

We should add this constraint to our extract_frames.pyscript.

• specifiy number of output samples wanted
• extract frames equidistant from video

Store frames/flow data in HDF5 format

We only have labels 0 or 1 for each class depending on its prediction. There are never going to be negative values. We might need to threshold on 0.5 since we probably get probabilities.

tmbo/caffe@fbe265a#diff-831821ebfa7329c6d3ccc472ad52626aR68

We need to split our data into train and test. Script should read the "filaname label".txt and split it according to a parameter into two such files.

We are in general very short on training examples. Therefore we should generate new ones based on existing ones:

• slight rotation
• move face patch slightly
• play with saturation / brightness

Currently we are generating a separate hdf5 db for x and for y flows. That doesn't make too much sense since we need them in one. They should be stacked in alternating order (x1, y1, x2, y2,...) to get a 20 channel blob.