Regarding the issue of training parameters about iadg

[splendor1811]

can you send me link to download O, C, I data please? my email is [email protected]

[almeidaraul]

Hello, did you manage to obtain closer results to the ones reported in the original paper? I am also facing very poor performance using the same code, or even using different (random) seeds.

[qianyuzqy]

@Elijah-Yi @almeidaraul Hi, sorry for the late reply. The training parameters are in the config files, e.g., https://github.com/qianyuzqy/IADG/blob/main/configs/ICM2O.yaml.

1. Do you use the same data processing and augmentation as our code?
2. Do you use the auxiliary supervision of depth estimator?

[Elijah-Yi]

@Elijah-Yi @almeidaraul Hi, sorry for the late reply. The training parameters are in the config files, e.g., https://github.com/qianyuzqy/IADG/blob/main/configs/ICM2O.yaml.

1. Do you use the same data processing and augmentation as our code?
2. Do you use the auxiliary supervision of depth estimator?

thanks for your reply

1. I used the same code parameters without modification, but changed the LMDB data reading section to the reading section. The txt writing method is set according to the image name, label, facial position, and 5 points, such as：
   “
   raw_dataset/FAS/image/Oulu_NPU/Train_files/6_3_05_4/88.jpg spoof 320.71435546875 345.529541015625 745.32421875 891.6963500976562 0.8942413330078125 445.04520124197006 538.2969526052475 636.3571692109108 558.3962008953094 540.9647116661072 646.6818950176239 451.160870552063 760.738587141037 609.6028578877449 771.9393855929375
   raw_dataset/FAS/image/Oulu_NPU/Train_files/6_3_05_4/93.jpg spoof 320.2257385253906 344.9953308105469 747.577880859375 893.0773315429688 0.8879839777946472 444.5309854745865 538.2054784297943 637.1285935640335 558.0140902996063 541.873034954071 646.9481581449509 449.9932810664177 759.6391240358353 609.7684855461121 771.432706952095
   raw_dataset/FAS/image/Oulu_NPU/Train_files/6_3_05_4/99.jpg spoof 321.75201416015625 345.2725524902344 747.1383666992188 892.3143310546875 0.8917432427406311 446.2833251953125 539.126389503479 639.0888638496399 557.0304822921753 543.0171566009521 646.5409145355225 451.85887384414673 760.4443521499634 612.8765425682068 771.5395402908325
   raw_dataset/FAS/image/Oulu_NPU/Train_files/6_3_05_4/104.jpg spoof 318.54931640625 344.02093505859375 745.1798095703125 893.1298828125 0.8957871794700623 442.6129494905472 536.9285832643509 637.1705750226974 552.4429047107697 539.9215207099915 644.3328658938408 449.3117601275444 756.1676215529442 611.2548422813416 767.8956568241119
   ”
2. Use PRNet for depth estimation, only perform depth estimation on positive samples

[Elijah-Yi]

Hello, did you manage to obtain closer results to the ones reported in the original paper? I am also facing very poor performance using the same code, or even using different (random) seeds.

I did not modify the experiment using the author's default code

[qianyuzqy]

Hi, @Elijah-Yi @almeidaraul ,
Sorry for the late reply.
Our work has been successfully reproduced by two recent works accepted to CVPR 2024 [1-3].
Moreover, the code of [3] is based on our codebase of IADG. You can have a look at this github repo.

Thanks for your interest in our work.

Best Regards,

[1]. Suppress and Rebalance: Towards Generalized Multi-Modal Face Anti-Spoofing, CVPR 2024
[2]. Gradient Alignment for Cross-Domain Face Anti-Spoofing, CVPR 2024
[3]. https://github.com/leminhbinh0209/CVPR24-FAS

[qianyuzqy]

@splendor1811
Sorry for the late reply. You can download the datasets and pre-process them following these links [1-2].

[1]. https://github.com/leminhbinh0209/CVPR24-FAS
[2]. https://github.com/sunyiyou/SAFAS