Nowadays, Artificial Intelligence is widely used in public spaces and this fact has raised a general concern about identity privacy. For this reason, face blurring has become a must-to-have feature for most applications to protect the identity. Unfortunately, in face blurring applications, the detection models used are not perfect, they can miss faces in some frames, which becomes a breach in privacy. For those cases, adding a tracker can be the solution to improve the system's robustness and enforce identity protection.
A common situation where detectors struggle is when the target face becomes partially obscured by another object or there is some level of rotation that makes it difficult to determine whether there is a face or not. In these cases, the face is still visible, but the detector may fail to recognize it. For instance, if a person’s face is covered momentarily by a hand or other object, the detector might lose track, even though the face is still present in the scene from a human eye perspective.
The following examples demonstrate the impact of using a tracker in a face detection application. The videos show two scenarios: the left side displays the result of using only the detector, while the right side shows the result when combining the detector with a tracker.
In the first video, featuring two young girls, we can see how the face of the short-haired girl is intermittently undetected by the detector (left side) when hands pass in front of her face. This leaves her face briefly exposed when it should be blurred. On the other hand, with the tracker enabled (right side), the face remains blurred continuously, demonstrating the tracker’s ability to maintain the censorship even when detection fails. When the tracker is used, it uses the information of the last detection available and extrapolates it to the next frames, so even if the detector fails on recognizing the face for some frames, the tracker covers it until the detector recovers.
Similarly, in the office scenario shown in Figure 2, we observe on the left (detector only) that the faces of two individuals on the left side are occasionally missed due to hand movements. In contrast, on the right side (with the tracker), these faces stay blurred throughout the video.
It is important to note that the tracker depends on the detector. Essentially, the tracker follows the object detected by the detector for a specified number of frames if the detector loses track of it. That’s why, in the office example, the person on the far right is not blurred in some frames on both the left and right sides, because the detector failed to detect the face from the start, meaning the tracker had no object to follow i.e. there is not a previous detection to track.
In conclusion, while trackers can significantly enhance face blurring systems, particularly in challenging situations where faces are partially obscured, they are not a replacement for accurate detectors. Instead, they complement detectors by maintaining object tracking for a certain number of frames, helping to minimize the impact of occasional detection failures and improve privacy.
Try It Out for Free!
If you are looking for adding more privacy to your video solution, look at our hardware-accelerated Deep Learning Face Blurring library. You can try it out for free at https://face-blurring.ridgerun.ai/.
Ready to Take Your Project to the Next Level? Contact Us!
We've shown how using trackers can significantly enhance object detection applications. Do you need help with your project or want to implement similar solutions? We’re here to assist you. Reach out to us at support@ridgerun.ai and let’s discuss how we can take your project to the next level together.
Comments