Skip to content

Latest commit

 

History

History
46 lines (32 loc) · 2.51 KB

Notes.md

File metadata and controls

46 lines (32 loc) · 2.51 KB
Raw

Dataset

refs(umd).p

The refs(umd).p file splits the dataset into three parts:

  • Train (img['split'] == 'train') - 42226 examples
  • Val (img['split'] == 'val') - 2573 examples
  • Test (img['split'] == 'test') - 5023 examples

For a total of 49822 examples.

Each object in refs(umd).p contains the following attribute: 'file_name': 'COCO_train2014_000000380440_491042.jpg'. It is not the actual image name. The actual image name is that obtained by removing the last part (e.g. _491042). The last part (the number following the last underscore) is the same as the attribute ann_id.

The attribute image_id uniquely identifies each image. There can be, indeed, more than one instance with the same image_id because the actual image is the same with different description/prompts/referring expressions.

Grouping images by image_id shows that there are 25799 different/distinct images.

Each image can have many sentences, that is prompts describing the content of the bounding box (note: the bounding box is instead located in the instances.json file). Computing the number of distinct images (that is, 25799, as shown above), computing the number of overall prompts, and dividing the latter by the former number gives ~3.7 (that is, there's an average of 3.7 referring expressions per image).

instances.json

It is a dictionary, made up of five top-level objects:

  • info
  • images
  • licences
  • annotations
  • categories

The images objects contains 25799 elements, therefore it is quite likely it contains a reference/description of each image.

Each object in images represents an image. It looks like they all share the same structure. They have 8 attributes, among which:

  • file_name: it seems the actual file name (not as that in refs(umd).p!)
  • id: it corresponds to the image_id attribute found in the objects loaded from the pickle

The attribute categories contains 80 elements, one per category. Each category has an associated id and a supercategory (which, in case of "base" categories, seems to correspond to the category name).

There are 208960 annotations. Each annotations contains:

  • segmentation
  • area
  • iscrowd
  • image_id: it corresponds to the image_id attribute found in the objects loaded from the pickle
  • bbox: the bounding box. It is an array of four values, likely to be [top_left_x, top_left_y, width, height]
  • category_id: it corresponds to the id attribute of the category
  • id: a unique id for the annotation (indeed, there are 208960 distinct ids)