In our lab, we tested the Multi Person Multi Camera Tracking TAPPAS pipeline using two cameras. While using it, we noticed some curiosities or inaccuracies with the models.
The model sometimes classifies two different people as the same person in a single camera feed, i.e. my colleague and I both get assigned the ID 1
The model seemingly “gets confused” when a person is visible on multiple camera feeds at once, often assigning two different IDs to the same person
We were wondering whether these inaccuracies stem purely from the model (then that’s that, not much we can do there), or if these arise based on the implementation, e.g. overlapping camera perspectives are not supposed to be used.
The model outputs an embedding vector that you can match to previously seen embeddings. This is an open-source model, so its accuracy is probably not state-of-the-art.
However, you can add additional heuristics to improve accuracy. The demo itself is not very complicated and uses simple heuristics. It utilizes our bbox-based tracker along with ReID embeddings.
There is no built-in verification of the “quality” of the images from which you obtain the embeddings. Additionally, the “threshold” for recognition can likely be fine-tuned.
You can achieve significantly higher accuracy by incorporating heuristics—for example, if the same person appears in two locations simultaneously, one of those detections is likely incorrect (neglecting quantum effects ). In such cases, you should prioritize the detection with the higher match accuracy.
Furthermore, you can leverage knowledge of camera locations to enable tracking across different camera sources. Another potential improvement is integrating face recognition, depending on the positioning and resolution of the cameras.
To summarize, there are many ways to improve accuracy, depending on how much effort you’re willing to invest.
). In such cases, you should prioritize the detection with the higher match accuracy.