NNID

We build what we name the Nearly-Nested Image Datasets (NNID) such that each dataset owns images of the same dimension, and each dataset is issued from a cropped version of the images belonging to the dataset with the biggest dimensions. This last dataset is named mother dataset and the images are named mother images.

By using NNID we ensure that the development is the same in all of the datasets. We also impose, as an additional constraint, that the difficulty of each dataset is the same. By the same difficulty, we mean that the distribution of costs is the same whatever the dataset. This additional constraint implies a specific way to crop the images, and most importantly, ensure that the experimental results obtained between the various dimension will be comparable since the source cost distribution of each dataset is the same. With the NNID we are able to avoid any impact of the development or the difficulty, on the experimental results; All the datasets are very similar except for the dimension.

We use, as a mother dataset, the LSSD dataset. LSSD is a mix of RAW images from ALASKA#2, BOSS, StegoApp DB, Wesaturate, RAISE, and Dresden datasets and uses a modified development script issued from the Alaska competition.

Given the NNID and an embedding algorithm, we focus on the average accuracy obtained by a classifier for each dataset. In order to obtain the relative payload size to embed for each dimension, we go by a dichotomous method, by running, for each dimension, multiple detections until finding the desired accuracy.