Resume
Well... In the last month, there were only few changes that came in this release, but before there was a lot of research and testing, to come up with this little release. You can see the PR #294 to see part of the main discussion.
But the main conclusion is that we discovered that model 2 has many grouping problems, and therefore we should advise against its use. 😞
If you are using it, sorry, but we suggest you migrate your faces to model 1, 3, if you dare, to model 4. However, I must recommend the analysis from scratch when switching to any model.
Perhaps the most important thing about this release is the introduction of model 4, which in summary is an extension of model 1, where we do a validation of the faces using the detector of model 3, to avoid clustering faces which potentially cause problems.
As you can see in this image, these are faces that the CNN detector finds, but the landmark detector fails to align the faces. We are fortunate that HOG does not detect these faces. and we take advantage of that, to validate the faces, and not group these faces. So, you can still see these faces, you can name them, but they will not clustered with others causing errors.
In our tests with 433 Images of The Big Bang Theory we got that:
433 Images | Model 1 | Model 2 | Model 3 | Model 4 |
---|---|---|---|---|
Faces | 1405 | 1405 | 1024 | 1405 |
Total clusters | 757 | 730 | 423 | 820 |
Faces in Bad Clusters | 13 | 63 | 0 | 0 |
Faces Error % | 0,93 % | 4,48 % | 0,00 % | 0,00 % |
As you can see, the quality of groupings improved dramatically. 😃
(and here we reaffirm avoiding model 2. Sorry again..)
...but this is also expensive:
433 Images | Model 1 | Model 2 | Model 3 | Model 4 |
---|---|---|---|---|
AVG Time | 16,70 | 16,59 | 41,56 | 50,21 |
Procesing Time | 02:00:29 | 01:59:43 | 04:59:56 | 06:02:19 |
Percentage | 100,00 % | 99,37 % | 248,93 % | 300,71 % |
It is two times slower. 😞
Well, It's still in testing stages, but it looks promising and everyone will know if it is worth it. 😉
Changelog
All notable changes to this project will be documented in this file.
[0.6.2] - 2020-08-14
Added
- Introduduce a new model (Model 4, aka DlibCnnHog5) that is 2 times slower, but
much more accurate, which now is in testing stages, and we invite you to test
since probably will be the next recommended model. See PR #313 for details. - Add face:migrate command that allows to migrate the faces obtained in a model
to a new one. Still recommended to fully analyze the images when changing
models, but can save a lot of time migrating them. See PR #309 - Add face:reset --model command to just reset current model.
Changed
- At least 1000 faces are needed for to make an initial clustering.
- Don't group faces smaller than 40 pixels, which are supposed to be of poor
quality. This is configurable within an advanced hidden setting. PR #299 - All reset commands require a confirmation to work.
- Hint the 4x3 relation when model recommending memory values.
Deprecated
- After many analysis, we discourage the use of model 2 (Aka DlibCnn68). We
still recommend model 1, and model 3 for low-resource devices. You can migrate
the faces using the new command, but we recommend analyzing them again.
Fixed
- Fix estimated time in the administration panel. See RP #297
- Fix that removing .nomedia file does not trigger facerecognition when next
analysis starts. Issue #304 - Fix travis tests and lot of scrutinizer reports.
- Fix that if increase the minimum confidence dont cluster any face in model 3
- Log the system info before return any error. Part of issue #278
Translations
- Add Macedonian translation thanks to Сашко Тодоров