2x False-Positives in 1 Snapshot. How to retrain faces seperatly and not together?

[ozett]

i have 2x wrong matches in 1 snapshot. i hopt that retraining manually helps the next time. but how do i retrain 2 faces in 1 images seperately [edit: from the gui]?

edit: i fetched the image from the /matches folder, cut out the faces and stored them manually in the correct folder for training. this seems the way without the webgui

[jakowenko]

There isn't a way from the UI to train an image with multiple faces. The detectors will reject the image because there's multiple faces in it, so the best way would be to manually crop them then upload them to training like you did.

Trying crop images with multiple faces and all of that from the UI is probably going to be a bit more than I want to support right now.

[Chreece]

There isn't a way from the UI to train an image with multiple faces. The detectors will reject the image because there's multiple faces in it, so the best way would be to manually crop them then upload them to training like you did.

Trying crop images with multiple faces and all of that from the UI is probably going to be a bit more than I want to support right now.

Or... What if we could crop the images direct in double-take's UI and train them direct? WIN-WIN (you David, loose a some of your free time :) )

[ozett]

i am on the latest betan and can already train single images from the gui.

a way to go would be to make the detection box clickable and save that face to the respective train-folder. but this happend for the first time, means seldom. so its ok to train manually..

[jakowenko]

Or... What if we could crop the images direct in double-take's UI and train them direct? WIN-WIN (you David, loose a some of your free time :) )

i am on the latest beta and can already train single images from the gui.

a way to go would be to make the detection box clickable and save that face to the respective train-folder. but this happened for the first time, means seldom. so its ok to train manually..

Both of those are good ideas. I can add it to the roadmap for a later release and prioritize it if there's more interest.

Personally my camera images usually aren't the best quality, so I rarely use them in training. I try to get high quality photos directly from my phone for the training.

[Chreece]

Or... What if we could crop the images direct in double-take's UI and train them direct? WIN-WIN (you David, loose a some of your free time :) )

i am on the latest beta and can already train single images from the gui. a way to go would be to make the detection box clickable and save that face to the respective train-folder. but this happened for the first time, means seldom. so its ok to train manually..

Both of those are good ideas. I can add it to the roadmap for a later release and prioritize it if there's more interest.

Personally my camera images usually aren't the best quality, so I rarely use them in training. I try to get high quality photos directly from my phone for the training.

I was training images through my camera, but since I read a post of yours suggesting only to train via good quality images (i.e. phone shots), I'm having better results! Thanks!

[ozett]

I was training images through my camera, but since I read a post of yours suggesting only to train via good quality images (i.e. phone shots), I'm having better results!

good to know how to improve matches. 👍 i will try with person i know and can take an closeup image must leave out faces that only came from the cams. hope to see differnences..

[jakowenko]

Started to work on a crop tool.

[ozett]

GREAT news!

[atv2016]

So why does a high quality image work better then a low quality one ? (really interested in how this works, as the actual image capture is going to be low res anyway?)

[bagobones]

I was just about to request a cropping tool for training images..

1. It is useful for cases like this where there are multiple faces
2. It can ensure that the model is training on an actual face not just some random thing it found in the photo.. Having seen how many times clothing patterns have had the face box (even with turning up the face certainty) I am not always sure the training photos are training what I want.. I would love to crop all the heads in all my training photos.

As a test I went and untrained a few people, took their photos into a photo editor and cropped them all down to face boxes and then retrained on the same data set.. I got VASTLY different detection results.. The result was better separation of people but overall lower certainty of matches which makes me suspicious of what it was locking on to in the uncropped photos.