Structure

After unzipping (no matter what version of the dataset you have downloaded) you will see the directory NS_AP_v1_0_a. Inside this directory, the examples are split into separate tasks (for full, and reduced dataset) or a single task (for task separated files).

A structure of the catalogue, given by one example, looks as following:

NS_AP_v1_0_a
| -- move_multi
|    | -- test
|    |    | -- images
|    |    |    | -- NS_AP_train_000400.png
|    |    |    ` -- ...
|    |    | -- scenes
|    |    |    | -- NS_AP_train_000400.json
|    |    |    ` -- ...
|    |    | -- sequences
|    |    |    | -- NS_AP_train_000400
|    |    |    |    | -- frame_0000.png
|    |    |    |    | -- ...
|    |    |    |    ` -- sequence.json
|    |    |    ` -- ...
|    |    ` -- instructions.json
|    ` -- train
|      -- val
` -- move_single
  -- move_weight
  -- pick_up_lightest
  -- stack
  -- stack_three
  -- stack_weight
  -- weight_multi
  -- weight_order
  -- weight_single

Each task has its separate folder with training, validation and test directories inside. Further, each of those contain images directory containing images generated for the initial state of the scene. Next, scenes directory contains .json files with scene graphs corresponding to the generated scenes. A file instructions.json contains a structure with instructions corresponding to the scenes in the present folder. Finally, a sequences directory contains folders for every example in that task and split. Those folders contain sequence.json with structure of observations collected while generating the example, along with rendered scenes in the form frame_XXXX.png.

Note that all imgaes, scenes, sequences share the same prefix, also indexing is preserved throughout whole split (e.g. move_multi contains indexes from 4000 to 4999 for training and 400 to 499 for validation and test).