Identity Problem (Different identity inputs towards the same identity output)

MJumbrella commented 4 months ago

Hi, thanks for the great work and I played with the demo provided. Specifically, I changed the input img_crop variables in the file demo_video.py to photos with different identities :

# 讀取範例圖片
img_crop = cv2.imread('./examples/0029_01_03_0022.jpg')
img_crop_2 = cv2.imread('./examples/0101_01_04_0018.jpg')
img_crop_3 = cv2.imread('./examples/0114_02_03_0011.jpg')
img_crop_4 = cv2.imread('./examples/0131_01_03_0014.jpg')

But when I ran the file, it gave me the output with the same or similar identities. Photos are attached for reference. Thanks for your help! identity problem

ms-ufad commented 3 months ago

一样的问题，自己的数据出来的还是测试用例的id

andreYoo commented 3 weeks ago

This issue seems to be caused by the style_encoder (Solver. extract).

IT IS VERY WEIRD THAT EVEN THO I PUT SOME FEMALE IMAGE, THE ENCODING SIMILARITY BETWEEN THE LATENT FEATURE EXTRACTED FROM THE FEMALE IMAGE AND THE OLD MAN (THE TWO SIMILAR OLD MEM IN THE ABOVE PIC) IS OVER THAN 0.95 (COMPUTER BASED ON COSINE SIMILARITY)

I am not sure that re-training the style encoder can help this issue because the entire training procedure for this framework has not been released yet.

I found out that the current style encoder encodes all images into the four identities.

So, whatever you put, the results will be aligned to one of the four identities provided by the project admin.

AvLab-CV / PASL

Identity Problem (Different identity inputs towards the same identity output) #1