search

bilylee / SiamFC-TensorFlow

A TensorFlow implementation of the SiamFC tracker
MIT License
358 stars 112 forks source link

关于data augmentation #96

Open tongtybj opened 4 years ago

tongtybj commented 4 years ago

请问这里的transformation为什么要 -8 ? https://github.com/bilylee/SiamFC-TensorFlow/blob/master/datasets/dataloader.py#L38

另外，因为下面的处理，instance embeds的shape永远都是[batch_size, 255-28, 255-28, 3]吗？ https://github.com/bilylee/SiamFC-TensorFlow/blob/master/datasets/dataloader.py#L39

tongtybj commented 4 years ago

Is it because the random stretch may resize the raw image with a -0.05: https://github.com/bilylee/SiamFC-TensorFlow/blob/f572dca95f2b3b2861f54de467259753428e468c/datasets/transforms.py#L49?

bilylee commented 4 years ago

这是为了对训练图像做 +- 4 个 pixel 的平移。如果直接 RandomCrop(255 - 2 * 8) 会导致平移 +- 8 个 pixel，所以先CenterCrop，再 RandomCrop。相关问题 #19 #32 #56