Summary
The paper proposes a new Face Recognition diffusion-based generation method. The diffusion process is completed with a semi-hard constraint on the synthetic reconstructed image: for each inquiry image of the (real) training set, the reconstructed image after the forward-backward diffusion process must have a specific cosine similarity with the inquiry image.
As it is usual for such methods in Face Recognition, the resulting synthetic dataset is then used for training a Face Recognition model. This model is evaluated across diverse real datasets.
Strengths
The tackled problem is quite hard and needed at the same time. Current SOTA Face Recognition generation methods lead to a significant gap in terms of performance, compared to real Face Recognition datasets (of the same size).
The idea of controlling the similarity to design semi-hard samples is also interesting.
Weaknesses
1) In Fig. 1, is the displayed similarity really the cosine similarity ? In the generated samples, the line with perfect similarity (equal to 1) seems to provide synthetic images which would not have a perfect similarity with the inquiry images displayed above the hypersphere.
2) [minor] In Eq. 3, the probability distribution of epsilon is not specified.
3) The authors should cite explicitly the works that use the training loss (Eq. 2) in this precise form, as there are alternative loss functions for diffusion models. A discussion on the reasons of this particular choice of diffusion loss might be a plus (e.g. in the appendix).
4) [minor] Although the lines 114-116 are accurate, they are misleading the reader. The widely known representation of Face Recognition embeddings is that they lie onto a hypersphere of dimension N, where each embedding is a point of the hypersphere. Those embeddings are clustered by identity on this sphere and the identity centers are roughly at the center of those clusters. The hypersphere mentionned in this paper is a hypersphere of dimension N-1, where the identity center is at the center of the sphere.
5) In lines 120-125, the authors should detail the range of similarities to the identity center, for each of the 5 splits of the CASIA training set. Only the average similarity of each split is specified.
6) [minor] Figure 3 should be a bit more explained than just its caption.
7) [major] Lines 155-162 are not well written and it is hard to understand how the margin m is used to guide the diffusion process. In particular, F_1 and F_2 are not defined, while some unused F is mentionned. C_sim seems to be a vector of unknown size. Also, the temporal guidance is too briefly described.
8) [minor] Some hyperparameters' values (alpha_t/beta_t, lambda) are not specified.
9) The right part of Fig. 4 displays two curves that do not have the same meaning for the x-axis. For AVG, the similarity is a constrained similarity (m) for training CemiFace (i.e. a similarity between a real inquiry image and a synthetic image). For CASIA, it is the similarity between one real image and its identity center (not a real image). To sum up, for AVG it is a similarity between 2 images, while for CASIA it is between 1 image and its identity center. Thus, comparing the two curves does not seem meaningful.
10) [major] The CosFace loss is used to train on synthetic datasets, while AdaFace is used to produce (identity-oriented) embeddings for the CemiFace training set generation. There should be only one model for both tasks, for fair comparisons. On Table 6, training on CASIA with AdaFace gives better results than with CosFace, so one could attribute the good performance of CemiFace to the fact that the authors used a stronger model (AdaFace) to generate the synthetic dataset than the model used to train on this dataset (CosFace). In addition, there should be a part studying the impact of this AdaFace choice (i.e. another loss), at least in the appendix.
11) The ROC curve on IJB-B/IJB-C for all synthetic methods of Table 6 would be a plus, as the accuracy is easily saturated, and not really used in industrial use-cases. Previous papers (related works) provide such ROC plots.
Questions
1) Could you explain the last sentence of Section 4.2.1 (lines 260-261) ?
2) In Section 4.2.2, why is the range of training m equal to [0,1] while the previous subsection concludes with an optimal range [-1,1] ?
Limitations
1) [major] The SimMat loss seems to be an interesting idea to lead towards m-similarity to the inquiry image, during training. But the derivation of the MSE loss (Eq. 3) assumes that the reconstructed image should be the inquiry image, and not a new image having a m-similarity with the inquiry image. I may be wrong here but I think that the diffusion loss of Eq. 3 is mathematically valid if the forward diffusion process is symetric to the reverse diffusion process, which is not the case here.
2) [major] In Section 4.2.1, there should be a part studying the difference between the required m and the estimated m post-training. That means that, for any required m, it is easy to compute the estimated m, i.e. the similarity between the resulting reconstructed image and the inquiry image. This estimated m must be quite different than the required m because Figure 4 shows that the best m is m=0, meaning that there is a 90 degrees angle between the inquiry image and the synthetized image. If m was truly equal to 0, the performance of the model would be very poor. So, there must have a difference between the required m and the real estimated m (post-training).
This is also reflected in the conclusion saying that the best setting for m is to train CemiFace with m randomly sampled from [-1,1]. A similarity truly equal to -1 would bring a model with an astonishingly poor performance.
** Update **
I have increased my score from 4 to 5.