ENH: be more selective with Van Hateren images.

guruucsd / DifferentialEncoding

Differential Encoding model

1 stars 3 forks source link

ENH: be more selective with Van Hateren images. #66

Closed bcipolli closed 8 years ago

bcipolli commented 8 years ago

I needed this, so I went ahead and coded it up.

A few changes:

Grab up to 4 image patches per image.
Make reversal an option (off by default)
Reject any image patch with std < 0.025, image mean < 0.05 or > 0.95 (pixels between 0 and 1), or some combo close to both thressholds.
Try up to 5 times to get a good patch; if you fail all 5 times, then warn the user (this alerted me to blank image imk00011.iml)

@vishaalprasad if you have time to take a look, please do!

bcipolli commented 8 years ago

Over runs, I don't get any really bad images. On 500 patches (250 train, 250 test), I get 7 skipped patches:

    (skipping vanhateren patch; bad mean,std combo)
    (skipping vanhateren patch; bad mean,std combo)
    (skipping vanhateren patch; 0.478 > mean_diff)
    (skipping vanhateren patch; 0.481 > mean_diff)
    (skipping vanhateren patch; bad mean,std combo)
    (skipping vanhateren patch; bad mean,std combo)
    (skipping vanhateren patch; 0.472 > mean_diff)

and the following images: