How to train/improve object recognition

[ThatQGuy]

What are some ways to improve object recognition or train the ai model to get better at recognizing objects?? Face recognition is great via Qmagie but object recognition is disappointing. Is there a way to train the model to get better at recognizing certain things?

If I want to make it get better at recognizing wine bottles in photos how would I do that?

[ThatQGuy]

bump

[Gvm77]

I would like to read responses to this post too. I probably don't know much about what the three AI engines do but I have the following concerns /observations:
1. QuMagie shows the faces that were recognised in photos in the properties of that photo. However, there is no indication that QuMagie recognises faces in many photos as being the same one. In other words, faces are being recognised in photos but QuMagie/Multimedia Console are not recognising them as belonging to the same person.
2. I haven't found a photo yet whose properties include an object that was recognised in it, let alone an object that appears in multiple photos.
3. No photo properties yet include location information even though the location info is included in the photo's metadata.
4. The AI engine detecting duplicate photos seems to have completed ok, but where are the details of the duplicates?
5. The throughput of the AI engines increased as more images were processed toward the end of the process. If this means the machine was 'learning' as it processed more and therefore it could process quicker, then will/how will this learning feedback to better recognition of photos processed early in the process. Do I need to re-initiate a second and/or third cycle of recognition?

Can forum members comment on ant or all of these observations please so that I can better understand what to expect of AI and the QNAP apps.
TIA ... Greg

[Theliel]

I would like to read responses to this post too. I probably don't know much about what the three AI engines do but I have the following concerns /observations:
1. QuMagie shows the faces that were recognised in photos in the properties of that photo. However, there is no indication that QuMagie recognises faces in many photos as being the same one. In other words, faces are being recognised in photos but QuMagie/Multimedia Console are not recognising them as belonging to the same person.
2. I haven't found a photo yet whose properties include an object that was recognised in it, let alone an object that appears in multiple photos.
3. No photo properties yet include location information even though the location info is included in the photo's metadata.
4. The AI engine detecting duplicate photos seems to have completed ok, but where are the details of the duplicates?
5. The throughput of the AI engines increased as more images were processed toward the end of the process. If this means the machine was 'learning' as it processed more and therefore it could process quicker, then will/how will this learning feedback to better recognition of photos processed early in the process. Do I need to re-initiate a second and/or third cycle of recognition?

Can forum members comment on ant or all of these observations please so that I can better understand what to expect of AI and the QNAP apps.
TIA ... Greg

1º. Facial recognition is complex, and there are different solutions. In this case, QNAP first performs clustering, grouping the images based on the similarity between them, according to different facial features. This grouping is done based on a pre-established threshold (which can be modified with a little knowledge). The higher the threshold, less faces have to be to fall into the same grouping but less false positive too; the lower the threshold, the greater the probability of false positives but more faces. Once the grouping is done, as you label the groupings and combine the ones that are really the same, the system can use them as training data, and be used to make inference in later recognition

2º. Object recognition uses a pre-trained model, which AICore passes to all the images you have in the library. These models are trained to identify a certain number of objects or themes in the photos. The photos are not labeled in this case, they are grouped into the different categories of "objects" that the model is able to identify. You can see all of this in QuMagie if you go to "Explore" and "Things". Possibly they are using Yolo or a similar model.

3º. I have a library of over 200K images, with a fairly high proportion of them with the location set, and virtually all of them can see the location. It is true that if we use the "Places" option within "Explore" in QuMagie, there is still a lot to improve, since the map view generally does not work due to a JS error, and only "Places" appear within a database that in my opinion is a bit ambiguous and meaningless, they could group it by cities at least, or localities. But when it comes to being able to see the location of the photos, there is no problem, the problem is that the maps are not visible due to code errors.

4º. I don't understand what you mean by details of duplicate photos. QuMagie has a "category" of similar photos, if you click there you can see them all, and each photo has information related to it

5º. It is answered in the first one. Without labeling anything, the system does not "learn", it does clustering to group the photos based on their similarity, obviously with a pre-established internal threshold. As you tag photos and group them, QuMagie can use that new data for better identification. Yes, ideally you could create a trained model with the photos already identified and grouped, and use it as a base model to make inference after clustering, but it is not something that can be done with QuMagie. This is what other systems like Digikam do. For example, I do not use QuMagie Facial Recognition for different reasons (they are identified but I do not pay attention to it), but mainly because through Digikam I have already recognized tens of thousands of them, and the recognition data is registered in the image's own metadata, which also allows it to be displayed by QuMagie as Tags. In this case Digikam works differently because it does not do Clustering, Digikam requires that you manually identify a certain number of photos that serve as training data to then make inference. As more faces are identified and labeled, the recognition model increases. At a certain point I can do two things, I could use that same model elsewhere, or since they are all already identified, train the model with the hundreds of thousands of people already identified.

[Gvm77]

Thank you very much Theliel. I will investigate your responses and get back to you with any questions if that is OK .... Greg