50,000 Test Images

Here is a set of 50,000 images from ImageNet that were constituted as a test set (not training set). The images, which are contained in the compressed tar files below, are in numpy format, and can be loaded into python as follows:

  import numpy as np
  d = np.load('1234.npy')   # This will load image index '1234' into 'd'

Each image is 227 x 227 pixels, and will load as an array that is 3 x 227 x 227. The RGB values in the images are very roughly in the range of -1.0 to +1.0, but they are not strictly limited to this range.

Below are links to ten compressed tar files, each of which has 5000 images. The linux command to unpack these files is, for example,

  tar xf i50k_p0.tar.z

Image files:

• i50k_p0 0 to 4,999
• i50k_p1 5,000 to 9,999
• i50k_p2 10,000 to 14,999
• i50k_p3 15,000 to 19,999
• i50k_p4 20,000 to 24,999
• i50k_p5 25,000 to 29,999
• i50k_p6 30,000 to 34,999
• i50k_p7 35,000 to 39,999
• i50k_p8 40,000 to 44,999
• i50k_p9 45,000 to 49,999