search

Ha0Tang / GestureGAN

[ACM MM 2018 Oral] GestureGAN for Hand Gesture-to-Gesture Translation in the Wild
http://disi.unitn.it/~hao.tang/project/GestureGAN.html
Other
174 stars 20 forks source link

Filter failure cases manually from train set? #6

Closed jysa01 closed 3 years ago

jysa01 commented 3 years ago

I am interested to understand the working of gestureGAN_twocycle model. I downloaded the senz3d dataset and prepared training and test data as indicated. When I saw the training output, I noticed that there are many images for which the pose is wrongly identified. Did you manually delete them from the 135,504 training samples? Can you please provide a .txt file with the correct sample's filenames ? Thanks and best regards, jysa01

Ha0Tang commented 3 years ago

Can you share with me some examples?

jysa01 commented 3 years ago

Below are some samples from the train set : G5_S3_7-color_G6_S3_1-color_AB_lr.png image G5_S3_5-color_G1_S3_13-color.png image G3_S4_20-color_G6_S4_1-color_AB_lr.png image

Ha0Tang commented 3 years ago

If you prepared the data according to the instructions I provided, there should be nothing wrong. I used all 135,504 samples for training.

jysa01 commented 3 years ago

OK, Thank you :) Yes, I did follow all the steps. From my understanding, the Matlab code provided prepares each data sample as [A,B,C,D] by reading each filename given in senz3d_split/senz3d_train.txt from the folders senz3d_image_skeleton/senz3d_skeleton and senz3d_image_skeleton/senz3d_image respectively.

So, the skeleton inside the senz3d_skeleton folder for the sample G5_S3_5-color_G1_S3_13-color.png is: G5_S3_5-color.png G5_S3_5-color

I seem to have interpreted this as a wrong pose. Thank you for the clarification.