README.md

---
license: apache-2.0
---

Here we create two datasets (from existing datasets: CLEVRER, VisualGenome) for the Object Counting instruction tuning task.

### CLEVRER, a video dataset

CLEVRER has QA pairs for each 5000 training videos.
```json
{'video_filename': int, 'scene_index': str (same as filename), 'questions': list [{'question_type': , 'question_subtype': , 'question_text': , 'answer_text': , 'program'(question attributes): }]} 
```
We select 'descriptive' type, 'count' subtype questions, they are object counting task questions. In the 'program' list, it shows how complex the question is (longer means more complex), so we filter out those longer than 9 to reduce difficulty.

CLEVRER contains both positive questions and negative (asking for non-exist objects) questions, so we skip generating negative samples for CLEVRER.

Some questions are 'event' specific, counting moving/stationary objects when a certain event happens. i.e., 'How many objects are stationary when the yellow object enters the scene?'

Downloading videos from: http://clevrer.csail.mit.edu/


### VisualGenome, an image dataset

We generate some negative questions for non-exist objects in the image. We use the version 1 image sets. Download from: https://homes.cs.washington.edu/~ranjay/visualgenome/api.html

VisualGenome has 100K+ images. And for the objects in the image, there are attributes associated with each object, we only focus on the color attributes.

For each image, we choose to add (1) 3 non-exist objects and (2) 1 non-exist attribute for existing objects as negative samples.

In the original qa dataset, VG has Object Counting questions, we also include them here, with the 'orig_qa'=='Yes'. For those negative questions we generated, 'orig_qa' =='No'.
```json
{'img_id': str, 'orig_qa': Yes/No, 'question_text': 'How many <attribute> <object in plural form> are there? ', 'answer_text': Numbers. or None. }
```
For more details, plz refer to the dataset.