MTCNN comes without Non-Maximum-Suppression (NMS)
When running our MTCNN face detector, we get a lot of overlapping detections. Typically, these are removed with a non-maximum-suppression algorithm, see for example here: https://github.com/TropComplique/mtcnn-pytorch/blob/45b34462fc995e6b8dbd17545b799e8c8a30026b/src/detector.py#L120 or in our TinyFaces implementation: https://gitlab.idiap.ch/bob/bob.bio.face/-/blob/de683894f9f14876293ad56390f4c34e7dd83234/src/bob/bio/face/annotator/tinyface.py#L229
However, our MTCNN implementation returns the outputs of the network unfiltered, leading to many overlapping detections: https://gitlab.idiap.ch/bob/bob.bio.face/-/blob/de683894f9f14876293ad56390f4c34e7dd83234/src/bob/bio/face/annotator/mtcnn.py#L113
When using only the first annotation as often done in our pipelines, this is not a big issue since NMS would just remove the overlapping boxes. When we need to detect more than one face in an image, on the other hand, we get a lot of repeated detections.
I would recommend to make the NMS function from TinyFaces accessible for other functions, and make use of it in MTCNN as well to filter out overlapping faces.