Version 0.9.19-beta.1240 is already available for download for Windows and macOS platforms
Major face recognition update
The previous version of Tonfotos’ facial recognition system performed well on adult faces, but had great difficulty with children’s and infants’ faces. In the new version, the clustering algorithm has been completely redesigned. In addition to the fast algorithm that was used earlier, another one was added, which is noticeably slower, but much more accurate. In addition, the new algorithm uses the date the photo was taken as a feature, which greatly simplifies the task of identifying babies.
Now Tonfotos combines the work of both algorithms. If for a person in the library there are still few examples of faces for training, a fast algorithm will be used, with which the user can quickly collect the necessary database of faces to train a more accurate algorithm. After that, the system switches to a more accurate one.
The results of the test show that now photos of children are classified as well as photos of adults. In addition, the algorithm copes much better with difficult cases - when the face is only partially visible, or when the person is photographed in profile, or even does not look into the frame at all. And also in low light conditions and low image resolution. As a result of the application of the new algorithm, the proportion of photographs of people that will be classified in semi-automatic mode has increased significantly, and now they are in a significant majority.
Also made a number of minor fixes and improvements in the recognition function, largely due to user feedback.
We invite users to evaluate the new algorithm and share their impressions.
New feature - “Offer less suggestions”
Since the new classification mechanism is quite powerful, it can be used to quickly mark up the majority of high-quality portraits for a person and run into a situation where there are no more normal photos left, and the proportion of correct faces among the proposed options begins to fall.
This is a normal phenomenon, if you have a library of 10,000 yet unidentified faces, and only 100 of them belong to the right person (and each time the number of correct ones decreases), the situation is inevitable when among the next batch of the most similar faces that the program offers, only a small part actually belongs to the right person. Nothing can be done about it - it’s pure statistics. This situation is a sure sign that you have already marked out most of the portraits of the person of interest. It is guaranteed that there are more in the collection, but finding them with each iteration will require more and more effort.
And here the user has a choice - to search further, to the bitter end, or stop. But it is difficult to stop if the system continues to suggest candidates. Yes, if you spend significant effort, then sooner or later you will give the system a sufficient number of “negative examples” that will help narrow down the search area and reduce the number of suggested faces. But not always there is time and desire to do this. Therefore, a new function has been added, just for this case. Through the context menu of a person in the people list, you can tell the program to start offering fewer suggestions for that person.
After that, the cut-off threshold for the proposed faces will be increased. They will continue to be offered, but only if confidence is high enough. Thus, the system will stop bothering you with “garbage” suggestions.
This action is fully reversible - at any time, also through the context menu, you can again lower the cutoff threshold and continue hunting for the last unallocated photos of this person.