Awesome

This the official repository of the Talk2Car-Expr dataset, an extension on the Talk2Car dataset, which is built on nuScenes. This dataset adds attributes and referring expressions to the referred objects in Talk2Car.

Data format

The training and validation data can be found in the data folder.

These files are dictionaries with the following data format:

{
"command_token":
  {"obj_box": [x, y, w, h],
   "class": "class_name",
   "img": "image.jpg",
   "action": action_value,
   "color": color_value,
   "location": location_value,
   "description": "referring expression"},

   ...
 }

The different values for actions, colors and locations can be found in data/vocabulary.json.

How to use

We provide a script, visualize.py, to show how to load the data and visualize it. You can use this dataset as a standalone or together with Talk2Car. We describe both options below.

Standalone

If you want to use this dataset as a standalone, please download the images as follows.

First, install gdown.

pip install gdown

Now download the images

gdown --id 1bhcdej7IFj5GqfvXGrHGPk2Knxe77pek

Unpack them,

unzip imgs.zip && mv imgs/ ./data/images
rm imgs.zip

Now the images will be stored in ./data/images

With Talk2Car

First, follow the instructions to install Talk2Car as described here. Then, copy the train and validation sets found in ./data to ./data/commands in Talk2Car. Now you can load the Talk2Car dataset with Talk2Car_Expr with the following:

ds = get_talk2car_class("./data", split="val", load_talk2car_expr=True)

This will load the Talk2Car_expr dataset. To access the data on the commands you can use the following:

# Print color of referred object of specific command
print(ds.commands[0].color)

# location
print(ds.commands[0].location)

# action
print(ds.commands[0].action)

Citation

If you use this data, please cite

@article{deruyttere2021giving,
  title={Giving commands to a self-driving car: How to deal with uncertain situations?},
  author={Deruyttere, Thierry and Milewski, Victor and Moens, Marie-Francine},
  journal={Engineering Applications of Artificial Intelligence},
  volume={103},
  pages={104257},
  year={2021},
  publisher={Elsevier}
}