[digiKam-users] 7.2 (beta) Face detection feedback

Sun Oct 25 21:03:12 GMT 2020

One more suggestion:

When clicking a person, it should be possible to toggle
confirmed/unconfirmed in the Thumbnails view.

Den søn. 25. okt. 2020 kl. 21.59 skrev Thomas D <sdktda at gmail.com>:

>
> I have a suggestion based on my previous:
> > Bonus: If there would be shown some kind of "confidence score"
> indicating how sure DK is that this is indeed the matched person.
>
> The more I use this face detection feature the more I think this is really
> helpful. I also think it would be really beneficial if DK would show the
> three or five best matches for any matches below a certain (high)
> threshold. For example, I have a hunch that DK *almost* detects correctly
> and maybe its second guess would be right. So then when I am going through
> Unconfirmed, I can quickly click one of these options. Of course I should
> still be able to write in the input field in case it is way off.
>
> What do you think?
>
>
>
> Den søn. 25. okt. 2020 kl. 20.59 skrev Thomas D <sdktda at gmail.com>:
>
>> I have one more:
>>
>> 11: When people have a long name that is too long to fit in the input
>> filed it appears that Digikam scrolls to the end of the name in the input
>> field. Example here:
>> [image: image.png]
>>
>> The problem is, that when you have lots of photos of people within the
>> same family they tend to all have the same last names. So it becomes
>> difficult to verify that DK detected the person correctly when showing only
>> the end of the name. And in order to see the first part of the name you
>> have to do this:
>> a) Hover the image with the mouse
>> b) Click the input field
>> c) Press Home button or press left arrow a bunch of times (taking several
>> seconds)
>>
>>
>>
>> I would suggest something like this instead: Each group presented is
>> instead various persons. And there could be a button to confirm the
>> detection for the entire group and a button to confirm  selected images in
>> that group.
>> Bonus: If there would be shown some kind of "confidence score" indicating
>> how sure DK is that this is indeed the matched person.
>>
>> [image: image.png]
>>
>> Den søn. 25. okt. 2020 kl. 20.27 skrev Thomas D <sdktda at gmail.com>:
>>
>>> Hi,
>>>
>>> Just want to say that I love digikam and use it extensively and have
>>> used it for over a decade. Keep up the good work!
>>>
>>> I also experience frequent crashes with face detection.
>>> Please let me know if there is anything I can do in order to help
>>> troubleshoot or debug this.
>>>
>>>
>>> I *think* I agree with the suggestion in #2 below. Although, I have read
>>> it twice and am not entirely sure I understand it.
>>>
>>> I have one addition: On my computer (modern 6 core i7 CPU, 32 GB RAM, 1
>>> TB NVMe SSD) DigiKam is almost unusable while "Detect faces" or "Recognize
>>> faces" processes are running. These processes run for a very long time.
>>> Thus, it would be really helpful if they did not influence the UI while
>>> running.
>>>
>>>
>>> Regarding #3: I am also hosting my photo collection on a NAS (samba
>>> share) and using DK on a Win10 client. I have the digikam database,
>>> thumbnail database, etc. stored locally on my C:-drive. I think this works
>>> OK. Can't you do the same in order to fix your problem? Or am I missing
>>> something?
>>>
>>> I would also be very interested in  some guidelines regarding #5.
>>>
>>>
>>> I have some more points here:
>>>
>>>
>>> (continuing numbering from previous list)
>>>
>>> 6: When going through Unconfirmed faces the UI is a bit confusing:
>>> there seems to be two buttons that does the exact same judging by the help
>>> text/mouse hover hint text. These buttons are seen in the screenshot here.
>>> The red X in the upper right cornor and the "minus button" at the bottom
>>> right. Both say: "If this is not a face, click to reject it".
>>> I think it is confusing that there are two buttons that does the same
>>> with very different symbols. Also I am in doubt: Do they actually do the
>>> same?
>>> If so, I think one should be left out. Also I am really missing an
>>> operation here. The operation I am missing a button for is this: Lets say
>>> that DK detected a face. This face is just some random person in the crowd
>>> of a concert or something. This is not a person I know or am interested in
>>> tagging. I do not want to "train" DK's ML model that it is not a face, so
>>> clicking "this is not a face" would be wrong. Because it IS a face. I just
>>> want to *ignore* this face as unintesting.
>>> I could, of course, create a person called UNKNOWN or John Doe or
>>> something like that and tag all faces like this as this pseudo-person.
>>> However, this is probably not a good idea, because then I will train DK's
>>> ML that this is actually a person that has many different faces. And I
>>> guess this will screw the ML up over time so that it will start to
>>> recognize a whole bunch of people as this unknown person.
>>> [image: image.png]
>>>
>>> 7: Another thing regarding the screenshot above. What does the minus
>>> button in the upper left corner do? In my version of DK, it does not seem
>>> to have any textual hint when hovering.
>>>
>>> 8: When tagging/confirming faces, it is a bit cumbersome to actually
>>> verify the face detection because you have to hover each image to see who
>>> DK detected it as. Instead, I think it should show the name directly on
>>> each face. This is almost the only relevant information to show except for
>>> the actual face image. I mean, stuff like file name, date, image
>>> resolution, size, etc. are much less important when doing face detection.
>>>
>>>
>>> 9: I have noticed that for some people I may only have a single photo or
>>> two of them. In these cases DK seems to find a LOT of spurious matches on
>>> these persons. Maybe this is something that could be tuned?
>>>
>>> 10: In my photo collection, I have photos of my children and other
>>> family members from back when they were babies and some of them are adult
>>> now. So I have *many* photos of the same persons in very different ages.
>>> However, DK seems to be confused by this. For example, when I have tagged a
>>> person as when they were baby, suddenly a lot of other babies become
>>> matched to this person. Is there any fix for this?
>>>
>>>
>>>
>>>
>>>
>>> Den søn. 25. okt. 2020 kl. 08.01 skrev <bryan at gillson.net>:
>>>
>>>> A few comments and questions on 7.2 beta and face detection:
>>>>
>>>>
>>>>
>>>>    1. I’m still suffering from occasional crashes, though they are
>>>>    more intermittent than 7.1. I know the Digikam team is already tracking
>>>>    these down as a priority, but let me know if logs would help.
>>>>    2. A suggestion for a UX change: After face detection, I
>>>>    multiselect thumbnails and confirm “unconfirmed” faces and manually tag
>>>>    others. Selecting the proper name causes the selected thumbnails to
>>>>    disappear, then reappear, then disappear (one at a time) as they’re
>>>>    processed. This shifts the remaining thumbnails around, making it difficult
>>>>    to multiselect another set of faces without error. I would prefer that
>>>>    faces “in process” simply stay hidden after selecting the name, and only
>>>>    appears again in the appropriate category (either “Confirmed” or in the
>>>>    proper People tag). This way I could continue “queueing up” new faces to
>>>>    process while KD works on the others.
>>>>    3. The face recognition workflow on a NAS-based setup really
>>>>    reinforces the need to separate thumbs.db into a local database.
>>>>    Performance is abysmal when refreshing thumbnails, and makes working over
>>>>    even a gigabit LAN very frustrating. Bug 297299 (
>>>>    https://bugs.kde.org/show_bug.cgi?id=297922) covers this feature
>>>>    request. Is it in the plan to address this?
>>>>    4. While on the topic of performance, are any bottlenecks being
>>>>    investigated to increase speed when using a network DB (MariaDB)? Neither
>>>>    CPU, network, database, nor disk statistics show high use during the
>>>>    recognition process. I would expect the database and network to be getting
>>>>    hit hard when retrieving pictures, then seeing the RAM and CPU spike during
>>>>    recognition. Instead, everything just goes at a trickle.
>>>>    5. Finally, is there any documentation on how to optimize the
>>>>    Detection process? What is more effective: confirming faces in the
>>>>    Unconfirmed tag? Correcting incorrectly recognized faces in a specific
>>>>    People tag? Adding new pics to known faces that are in Unknown, to better
>>>>    train the AI? Is it better to tag many samples of a person’s face, or a
>>>>    few? How frequently should I re-run “Recognize faces” after
>>>>    adding/correcting/confirming faces? Given the significant time investment
>>>>    in tagging and recognizing, guidance on the above could save dozens of
>>>>    hours of effort.
>>>>
>>>>
>>>>
>>>> I love Digikam, and truly appreciate all the effort. Please don’t take
>>>> the above as criticism or complaints, but questions/suggestions to make a
>>>> great product better.
>>>>
>>>>
>>>>
>>>> Thank you.
>>>>
>>>
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://mail.kde.org/pipermail/digikam-users/attachments/20201025/9c3066bb/attachment-0001.htm>
-------------- next part --------------
A non-text attachment was scrubbed...
Name: image.png
Type: image/png
Size: 9857 bytes
Desc: not available
URL: <http://mail.kde.org/pipermail/digikam-users/attachments/20201025/9c3066bb/attachment-0003.png>
-------------- next part --------------
A non-text attachment was scrubbed...
Name: image.png
Type: image/png
Size: 4095 bytes
Desc: not available
URL: <http://mail.kde.org/pipermail/digikam-users/attachments/20201025/9c3066bb/attachment-0004.png>
-------------- next part --------------
A non-text attachment was scrubbed...
Name: image.png
Type: image/png
Size: 50434 bytes
Desc: not available
URL: <http://mail.kde.org/pipermail/digikam-users/attachments/20201025/9c3066bb/attachment-0005.png>