What is the difference between LMDB and raw images

[정윤희]

I did fine tuning with VGG networks with both raw images and LMDB file.
But each of the results was very different. (result of LMDB was even better then raw images') although the networks was quite same apart from only input layer.
I don't know what the difference is between when I use LMDB and when I use raw images.

=== Raw images data input layer ====

  image_data_param {
     source: "data/face_finetune/train.txt"
    batch_size: 50
    new_height: 224
    new_width: 224
  }



===== LMDB data input layer ======
 data_param {
     source: "data/face_finetune/face_reid_train_lmdb"
     batch_size: 50
     backend: LMDB
 }

[jdh...@gmail.com]

The lmdb files also contain the image informations, but in lmdb format, which can be more efficiently processed by Caffe. If you directly process image input

using caffe, it will be slower. Apart from effiecy, I don't think the two  ways have any difference.

在 2016年3月31日星期四 UTC+8下午1:39:03，정윤희写道：

[gwang been park]

Now i faced with exactly same problem with you. When i tested my trained network with lmdb, accuracy is 98%, but when i tested same trained network with raw jpg files(my manual classification api is from example/cpp_classification.cpp code) it gives 95% accuracy that is slightly less than lmdb's one.... you solved problem? If it is, please tell me how to solve it and reason of this problem..