metadata

dataset_info:
  features:
    - name: image
      dtype: image
    - name: size
      dtype: string
    - name: objects
      dtype: string
    - name: positive_captions
      dtype: string
    - name: negative_captions
      dtype: string
    - name: ambiguous_captions
      dtype: string
    - name: positive_objects
      dtype: string
    - name: negative_objects
      dtype: string
    - name: ambiguous_objects
      dtype: string
  splits:
    - name: train
      num_bytes: 6424085843.442
      num_examples: 38118
    - name: test
      num_bytes: 1683318047.872
      num_examples: 9658
  download_size: 8236094882
  dataset_size: 8107403891.314
configs:
  - config_name: default
    data_files:
      - split: train
        path: data/train-*
      - split: test
        path: data/test-*
license: mit
task_categories:
  - object-detection
  - image-feature-extraction
  - image-to-text
language:
  - en
pretty_name: HICO-DET Dataset
size_categories:
  - 10K<n<100K

Dataset Card for HICO-DET Dataset

Dataset Summary

HICO-DET is a dataset for detecting human-object interactions (HOI) in images. It contains 47,776 images (38,118 in train set and 9,658 in test set), 600 HOI categories constructed by 80 object categories and 117 verb classes. HICO-DET provides more than 150k annotated human-object pairs. V-COCO provides 10,346 images (2,533 for training, 2,867 for validating and 4,946 for testing) and 16,199 person instances. Each person has annotations for 29 action categories and there are no interaction labels including objects.

47,776 images (38,118 in train set, 9,658 in test set)
600 HOI categories
80 object categories
117 verb classes
Over 150,000 annotated human-object pairs

Originally hosted at http://www-personal.umich.edu/~ywchao/hico/, the dataset is no longer available at its official website. This Hugging Face version is a converted and restructured copy of the original dataset, which can still be found on Google Drive in MATLAB format.

Dataset Structure

The dataset is structured as follows:

DatasetDict({
    train: Dataset({
        features: ['image', 'size', 'objects', 'positive_captions', 'negative_captions', 'ambiguous_captions', 'positive_objects', 'negative_objects', 'ambiguous_objects'],
        num_rows: 38118
    })
    test: Dataset({
        features: ['image', 'size', 'objects', 'positive_captions', 'negative_captions', 'ambiguous_captions', 'positive_objects', 'negative_objects', 'ambiguous_objects'],
        num_rows: 9658
    })
})

Here is the description of each column:

image: the image
size: the size of the image
objects: the object categories in the image
positive_captions: the positive captions for the image, e.g., ('cake', 'carry') which means the image is a picture of a person carrying a cake. One image can have multiple positive captions.
negative_captions: the negative captions for the image. One image can have multiple negative captions.
ambiguous_captions: the ambiguous captions for the image. One image can have multiple ambiguous captions.
positive_objects: the positive objects for the image. Positive objects are the index of (object, verb) pairs in list_action.csv.
negative_objects: the negative objects for the image. Negative objects are the index of (object, verb) pairs in list_action.csv.
ambiguous_objects: the ambiguous objects for the image. Ambiguous objects are the index of (object, verb) pairs in list_action.csv.

list_action.csv is a csv file that contains the list of (object, verb) pairs and some other useful information, which can be found here.

Usage

The most simple usage is to load the dataset with hugging face datasets.

from datasets import load_dataset

dataset = load_dataset("zhimeng/hico_det")
print(dataset)

File Structure

I also provide the original data structure as the following:

data/
    list_action.csv # 600 HOI categories
    images/
        train/
            metadata.json  # 38,118 images
            img_00001.jpg
            img_00002.jpg
            ...
        test/
            metadata.json  # 9,658 images
            img_00001.jpg
            img_00002.jpg
            ...