Skip to content

data_gen_inst.md

Latest commit

 

History

History
72 lines (53 loc) · 2.35 KB

data_gen_inst.md

File metadata and controls

72 lines (53 loc) · 2.35 KB
Raw

Data Generation Instructions

Hold vs. Non-hold

  • if the sign involves a gesture it is a non-hold
  • if the sign is simply a hand shape it is a hold

Where to store the videos

Store videos in a directory named with the label of the signs.

Confirm the quality of the videos

At any point videos can be rewatched with a mp overlay to ensure the tracking is good using the following holistic.py command

python holistic.py multi <video dir> <start index>

  • note this will create a folder in the video directory called mp with your videos with the mp overlay
  • it will also display the tracking as it happens, press q on the video window to skip to the next vid.

for example if I want to verify my current vids are tracking well for a sign apple I would use python holistic.py multi C:\signs\apple 0

or to start at the 10th video I would do

python holistic.py multi C:\signs\apple 9

Recording

hold videos should be done as follows:

  • length in multiples of 10 + 1 second for setup at start to get mediapipe to recognize the hand
  • data can be in as many videos as needed as long as the total recorded length is >240 seconds
  • in past I have done many 11 second videos with a few longer videos once I'm certain the gesture I'm making is tracking well in mediapipe

non-hold videos as follows:

  • 1 video per sign iteration
  • keep videos at or below 2 seconds including setup time
  • in past I have recorded a minimum 60 vids for these signs, but the more the better

store vids in the following structure

-> videos
    -> apple
        -> vid1.mp4
        -> vid2.mp4
        -> ...
    -> orange
    -> etc...

Data Generation

once vids are recorded generate data using the following holistic.py command

python holistic.py dataset <dir with sign subdirs> <output directory>

using the structure above, the following command: python holistic.py dataset /path/to/videos /output

will generate an output structure like this:

-> output
    -> apple
        -> vid1.npy
        -> vid2.npy
        -> ...
    -> orange
    -> etc...

this is the data, it can then be sorted into the hold and non hold directory

Committing Your Changes

Please make the changes in a new branch with the following naming convention: data/<signs separated by '-'>. Only commit the generated numpy data, DO NOT include the videos!