Kaggle-CMI

Repository for Kaggle's competition "IEEE's Signal Processing Society - Camera Model Identification".

1. Data

Train

• Images in the training set were captured with 10 different camera models, a single device per model, with 275 full images from each device.
• The pixels of each image are different for each device (same device, same pixels).

Test Images

• Pictures in the test set were captured with the same 10 camera models, but using a second device.
• While the train data includes full images, the test data contains only single 512 x 512 pixel blocks cropped from the center of a single image taken with the device.
• No two image blocks come from the same original image.
• Half of the images in the test set have been altered. The image names indicate whether or not they were manipulated (_manip) from the original or unaltered (_unalt).
• Operations
  • JPEG compression with quality factor = 70
  • JPEG compression with quality factor = 90
  • resizing (via bicubic interpolation) by a factor of 0.5
  • resizing (via bicubic interpolation) by a factor of 0.8
  • resizing (via bicubic interpolation) by a factor of 1.5
  • resizing (via bicubic interpolation) by a factor of 2.0
  • gamma correction using gamma = 0.8
  • gamma correction using gamma = 1.2

Additional Data (by Gleb Posobin)

flickr_images: Pictures taken with the phones from the dataset. There is also a good_jpgs file — that’s a list of all photos with resolution matching the ones in the provided training dataset.

Note!: Picturestaken with moto x that flickr provides are made with an old (2013) model, but for our task a 2014 version is used, so these photos should be discarded.

In total there are 5377 photos, after leaving only the ones that are in good_jpgs file.