Download the datasets

Manual download

Start by creating a LOOK folder.

Annotations

You can download the LOOK annotation file here and copy it inside the LOOK folder.

Nuscenes

Go to Nuscenes official website and download the samples and the CAM_BACK_LEFT sweeps from the server closest to your location (US or ASIA). Create a Nuscenesfolder inside the LOOK folder and extract the downloaded files there.

JRDB

Create a folder JRDB inside LOOK. Download the JRDB train dataset from here and extract it in the JRDB folder you just created.

Kitti

Inside the LOOKfolder, create a Kitti folder.

Training data

Create an account at the Kitti Benchmark website and move to the raw data section. Download the 2011_09_29_drive_0071 [synced+rectified data] folder, you can directly use this link.

Unzip the folder and copy the 2011_09_29/2011_09_29_drive_0071_sync/image_03/data folder into the Kitti folder you created. Rename it train.

Testing data

Move to the 2d Object detection benchmark here. Download the first folder called Download left color images of object data set (12 GB) or directly using this link. Unzip the folder and copy the folder data_object_image_2/training/image_2 into the Kitti folder. Rename it test.

Filter the dataset

To keep only the files with a LOOK annotation, you neeed to filter the datasets you previously downloaded. To do so, you can download the extract_look_dataset.py python script. This script allows you to extract only the annotated images and reorganize them in order the respect the paths of the LOOK dataset. To run this script you need to use python 3 running and the pandas library.
To be able to run the script you need to input the name of the dataset(s) you want to filter and the path to the annotation file. To run the script for all the datasets, you can navigate inside the LOOK folder in a terminal window and run the command
python extract_look_dataset.py -d Nuscenes,Kitti,JRDB -a LOOK_annotations.csv.

The dataset

Dataset	Frames	Instances	Looking	Pedestrians
nuScenes	2,216	13K	9%	7,100
Kitti	1,391	4,630	17%	425
JRDB	9,441	39K	18%	399
LOOK	13K	57K	16%	7,944

The LOOK dataset is made of images from 3 different existing datasets where we annotated the pedestrians as looking (1) or not (0) at the camera.

• • nuScenes contains 1,600x900 .jpg images.


• • Kitti contains .png images at different resolutions.


• • JRDB contains 752x480 .jpg images.