Pairwise Ranking Distillation for Deep Face Recognition

Abstract

This work addresses the problem of knowledge distillation for deep face recognition task. Knowledge distillation technique is known to be an effective way of model compression, which implies transferring of the knowledge from high-capacity teacher to a lightweight student. The knowledge and the way how it is distilled can be defined in different ways depending on the problem where the technique is applied. Considering the fact that face recognition is a typical metric learning task, we propose to perform knowledge distillation on a score-level. Specifically, for any pair of matching scores computed by teacher, our method forces student to have the same order for the corresponding matching scores. We evaluate proposed pairwise ranking distillation (PWR) approach using several face recognition benchmarks for both face verification and face identification scenarios. Experimental results show that PWR not only can improve over the baseline method by a large margin, but also outperforms other score-level distillation approaches.

Modular Multi-Face Tracking Geared Toward Face Recognition in Surveillance Videos

“…In the LTFT system, three scores are employed to assess the quality of each detected face: detection confidence, head angles and sharpness measure. The detection confidence is a value in the interval of [0, 1], given by the face detection model; the head angles are used to measure pose, being obtained through the use of the head pose estimation model 3DDFA [35] as a set of three values that represent the angular rotation of a face along the three dimensions axis; the sharpness is calculated as the average of the modified Laplacian filter's [36] response. The Table I below, illustrates four face images and their respective quality scores.…”

“…The employed sharpness measure assigns very small values and with little variation for the majority of detected faces (lesser than 0.1, these values are expected to be distributed between [0.0, 1.0]), we argue that this hinders the reconnection process performed by the face-based tracklet reconnection module. Thus, we propose to replace the modified Laplacian filter [36] by the method presented in [4] and described by Equation ( 4), where X i is the i-th face image detected on video, Sh Xi is the final sharpness score of image X i , avg is the average of all pixels, abs means absolute value and lowpass is a simple mean filter of kernel size 3x3. The modified Laplacian filter [36] requires the location of fiducial points along the detected face in an attempt to crop the face region before sharpness evaluation.…”

“…Thus, we propose to replace the modified Laplacian filter [36] by the method presented in [4] and described by Equation ( 4), where X i is the i-th face image detected on video, Sh Xi is the final sharpness score of image X i , avg is the average of all pixels, abs means absolute value and lowpass is a simple mean filter of kernel size 3x3. The modified Laplacian filter [36] requires the location of fiducial points along the detected face in an attempt to crop the face region before sharpness evaluation. Therefore, this replacement also has the added benefit of not necessitating a model for fiducial points detection.…”

“…• (I) LTFT: the LTFT system in its baseline state, using the Faceboxes model as face detector, original parameters and with T max = 5; • (II) LTFT+thresholds: same as the previous baseline experiment, this time changing the sharpness measure's thresholds, where n E = 0.06 and n V = 0.04; • (III) LTFT+Yolo5s: this experiment evaluates the contribution of changing the face detection model, where n E = 0.06, n V = 0.04 and T max = 5. • (IV) LTFT+Yolo5s+FQA: in addition to replacing the face detection model, the modified Laplacian filter by [36] is replaced by the face sharpness measure presented by [4] and the face image resolution is added to the FBTR module as an extra face quality assessment measure. • SORT: the SORT tracker applied with the Yolo5s detections.…”