This repository contains CSVs that list classes of varying levels of granularity and the mappings between these levels for the Something-something dataset.
Download the fine-grained labels from the Something-something dataset website
something-something-v2-train.jsonsomething-something-v2-validation.jsonsomething-something-v2-test.json, these just contain IDs for test videos.something-something-v2-labels.json, this is a mapping between (cleaned) fine-grained class names and numeric indices.
- Fine-grained classes (mid-level): the original 174 classes released with the dataset as
specified in
something-something-v2-labels.json. - Coarse-grained classes (high-level): A remapping of the fine-grained classes into 40 classes.
- Captions (low level): The fine-grained classes are actually templates. Templates contain 0 or more placeholders (surrounded by square brackets) for which the corresponding values are available for each instance.
These are the original 174 classes first described in the dataset paper.
These are stored in the template field in the JSON files provided by 20BN.
There are 50 high level classes derived from all the fine-grained labels.
fine_to_coarse.csv contains the correspondence between the 174 fine-grained
classes to the 50 coarse-grained classes.
These have to be mapped from the template field in the JSON files provided by
20BN in combination with fine_to_coarse.csv.
The fine-grained labels are actually templates parameterised with noun phrases, for example, here is the metadata for video id 207125:
- Template: Putting [something] onto [something else that cannot support it] so it falls down
- Place-holders:
- ice cream container
- water bottle
- Caption: "putting ice cream container onto water bottle so it falls down"
The template corresponds to the fine-grained label, and the placeholders combined with this produce the caption.
In total there are 123,858 unique captions in the combined training and testing split (9/10th of the dataset). Presumably the test set contains a reasonable proportion more unique captions.
These are stored in the label field in the JSON files provided by 20BN. The
original template is stored in the template field and the values that have been
substituted in are stored in the placeholders field.
The original paper describes a 10 coarse-class subset of the dataset where they
defined new coarse-classes (different from those described before). The 10 class
names are defined in 10_class_subset.csv and the mapping between fine-classes
and these is defined in fine_to_10_classes.csv.
A 40 class subset is described in the paper and is produced by combining the 10
class subset plus another 30 fine-grained classes. The class names are defined
in 40_class_subset.csv in addition to the corresponding fine-grained class
index to lookup in fine_grained_classes.csv and the coarse-grained class index
to lookup in 10_class_subset.csv.
Dataset Paper Table Reference
- Table 5:
10_class_subset.csv - Table 6:
40_class_subset.csv - Table 7:
fine_to_10_classes.csv - Table 8:
fine_to_coarse.csv
The coarse-grained classes in 10_class_subset.csv, 40_class_subset.csv, and
fine_to_10_classes.csv are DIFFERENT to the coarse-grained classes in
fine_to_coarse.csv and coarse_grained_classes.csv. The paper defined two
sets of coarse-grained classes for the 10/40 class subsets and the full dataset.
The data contained in this repository is entirely derived from either the papers below, or from the dataset itself.
- The original dataset paper: The "something something" video database for learning and evaluating visual common sense.
- A follow up paper using v2 of the dataset and the varying levels of labels: "On The Effectiveness of Task Granularity for Transfer Learning"
Many thanks to 20BN for providing such an interesting dataset to work on.