Evaluation

[3068615155]

May I ask how the evaluation metrics are computed? Is there any available code for this? Also, why do I obtain relatively low metric values when evaluating with the provided pretrained model? Thank you. @seungjun-moon @JH-LEE-KR @hahminlew
In addition, did the comparative experiment with GSAvatars also use the dataset provided on the website? I sincerely hope you can help clarify my confusion. Thank you again.

[hahminlew]

@3068615155 Hi, thank you for your interest in our work.

We computed MSE, PSNR, SSIM, and LPIPS by applying the same GT foreground masks, generated in the manner described in the paper (see Appendix 8.6), to the outputs of all baselines. Therefore, one possible reason for relatively low metric values is that the background may have been included during evaluation. For example, some baselines generate images with white backgrounds, some with black backgrounds, and others do not remove the background at all.

For the main table in the paper, all baselines were newly trained and evaluated on the single-view datasets (SplattingAvatar and DynamicFace) for each individual subject.

In addition, if by GSAvatars you mean GaussianAvatars, please note that it is originally a multi-view avatar generation model. For this reason, we reported an additional comparison in Table 2. In that experiment, we used the Nersemble dataset from GaussianAvatars, where GSAvatars was trained in the multi-view setting and GeoAvatar was trained in the one-view setting for quantitative comparison.

I hope this clarifies your questions. Please let us know if anything is still unclear.