YOLO dataset for WRS 2018 (world robot summit - partner robot challenge in real space)
We participated in WRS 2018 and we used YOLO to recognize target objects. This dataset is used there and used for training the YOLO v3. The objects can be shown here. The detailed labels of each object are as below and some objects which seem like that HSR cannot grasp are not annotated.
0 Dog plush toy
1 Rabbit plush toy
2 Hedgehog plush toy
Postcard with moon image
Postcard with temple image
Postcard with waterfall image
3 Bus toy car
4 Fire engine toy car
5 Police toy car
6 Whistle sound toy
7 Shaker sound toy
8 Horn sound toy
9 Cube block
10 Triangular block
11 Cylinder block
12 Wooden Bowl
Brown pail
13 Mixed nuts
14 Blue cup
15 Pink cup
16 Orange biscuits
17 Yellow biscuits
Watering can
18 Flashlight
19 Blue fork
20 Green spoon
21 Gray knife
22 Pink biscuits
23 Yellow clock
24 Aluminum foil
25 Green dish
26 Yellow dish
27 Pink bowl
28 Blue bowl
29 Mineral water
30 Oolong Tea
31 Ketchup
32 Canned mustard
33 Orange drink
34 Green drink
35 Tomato
36 Potato
37 Plant
38 Spray
39 Water bottle
40 Blueberry drink
41 Eggplant
The structure of each dataset is as follows.
|- Dataset
|--- Dataset 1 (real images)
|------- JPEGImages
|---------- xxxxx.jpg(or other image extensions)
|------- Labels
|---------- xxxxx.txt
|
|--- Dataset 2 (synthesized images)
|------- JPEGImages
|---------- xxxxx.jpg(or other image extensions)
|------- Labels
|---------- xxxxx.txt
The dataset have 2 datasets. One is composed of real images and the other is composed of synthesize images. Annotaion style follows the annotation style of YOLO. Therefore, you can use it directly for training YOLO.
Real images are taken by several cameras such as phone cameras, RealSense D435, and Xtion. The objects are placed at random sites in the demonstration place.
Synthesized images are generated by combining a background image and several object images. The background image is selected from ~100 background image candidates which collected by web crawling and taken around the target envrionment. The object images are taken in a studio and the background is eliminated. These object images are taken for all objects.
Dataset link: https://koreaoffice-my.sharepoint.com/:u:/g/personal/jhj0630_korea_edu/EZTuVivyBiBJn2WTiF7kRZ4Bm28TesN1DEmmKfx67r4idw?e=6xSvKA
This dataset is published under the LGPL license.