search

Ning-Ding / Implementation-CVPR2015-CNN-for-ReID

Implementation for CVPR 2015 Paper: "An Improved Deep Learning Architecture for Person Re-Identification".
MIT License
147 stars 71 forks source link

Data augmentation #19

Open MarcosPieras opened 7 years ago

MarcosPieras commented 7 years ago

Hi, First of all, thanks for your work!

In a previous issue you said that you have 130000 images, they are image pairs, are they ? In data augmentation, I understand that you double the dataset by a random transform.

Do you consider to add more transformation to enlarge your data set ? or it may be to noisy ? thanks!

Ning-Ding commented 7 years ago

Yes, they are image pairs without data augmentation. In the paper, the author said they made five random transform for each images. Also, they have flipped the image once. So after data augmentation, they have 10 images for one original images. Generally, data augmentation is very important for getting a good result. I would prefer a online random transformation during training, not a offline image preparation.