Oktoberfest Food Dataset

The data was aquired during Schanzer Almfest at Ingolstadt in 2018 by IlassAG. As part of a practical at the Data Mining and Analytics Chair by Prof. Günnemann at TUM we were given the task to count objects at checkout. Therefore we annotated the data with bounding boxes and classes to train an object detection network.

Download

You can find the dataset here

• dataset contains the train and test datasets including the labels
  • the labels can be found in files.txt (OpenCV style)
  • <filename> <number of objects> <classid1> <x1> <y1> <w1> <h1> <classid2> <x2> <y2> <w2> <h2> ...
• models contains our pretrained tensorflow models (see Preview.ipynb for an example usage)
• video_data_zipped contains the raw videos from which the dataset were extracted

The dataset in the PASCAL_VOC format can be found here

https://drive.google.com/open?id=1rgJUEFB4Cmbf9mQVdGPCHGiT4bvh_gDT (images)

https://drive.google.com/open?id=1mLIc1Ybs1rVwzMDuWMwxWUgl7spx2tBB (video)

In addition, the labels in the PASCAL_VOC format are available in the PASCAL_VOC folder.

Online Notebooks to train Faster RCNN and Retinanet models on the dataset using Google Colaboratory are available here

Faster RCNN Pytorch

https://drive.google.com/open?id=1CDQ5cIA8qsdm-OinbfPKM5DuoI6ewvZH

RetinaNet Tensorflow

https://drive.google.com/open?id=1KxP-j0TSQ_PY7xkJ4JNRyMnLv7kjRB_e

Dataset Description

Data Distribution

Class Id	Class	Images	Annotations	average quantity
0	Bier	300	436	1.45
1	Bier Mass	200	299	1.50
2	Weissbier	229	298	1.30
3	Cola	165	210	1.27
4	Wasser	198	284	1.43
5	Curry-Wurst	120	159	1.32
6	Weisswein	81	105	1.30
7	A-Schorle	90	98	1.09
8	Jaegermeister	43	152	3.53
9	Pommes	110	126	1.15
10	Burger	105	122	1.16
11	Williamsbirne	50	121	2.42
12	Alm-Breze	100	114	1.14
13	Brotzeitkorb	65	72	1.11
14	Kaesespaetzle	92	100	1.09
	Total	1110	2696	2.43

Statistics

Benchmark

In order to train object detection models, we used Tensorflow Object Detection API and standalone Pytorch models. We trained several different models and got the best results for the Single Shot Detector with Feature Pyramid Networks (RetinaNet). Our evaluation metric was the area under the precision-recall curve (AUC) on a test set of 85 images (we ignored localization as our goal was counting objects). We also used the data from the last two days to evaluate the models trained on the data from the first 9 days.

Approach	Backbone model	AUC	Example precision@recall
SSD	Mobilenet	0.86 / -	0.85@0.70
SSD + FPN	Mobilenet	0.98/0.92	0.97@0.97 (0.88@0.90)
RFCN	ResNet-101	0.97/0.89	0.90@0.95 (0.72@0.90)
Faster RCNN	VGG-16	0.98/0.93	0.97@0.90 (0.79@0.90)

The numbers on the left denote the result on the small test set, the ones on the right - on the test set of data for the last two days (819 images for training, 222 images for testing).

Code

The Evaluation folder contains Jupyter notebooks to evaluate the TensorFlow models.

With the Preview notebook one can try out the pretrained TensorFlow models on arbitrary images.

The CreateTFRecordFile notebook contains code to convert the dataset in to the TFRecord file format so it can be used with the TensorFlow object detection library.

The ShowAnnotations visualizes the bounding boxes of the dataset. Use 'n' for the next image, 'p' for the previous and 'q' to quit.

Authors

Vitalii Rusinov: Student of Informatics (M.Sc.) at TUM
Alexander Ziller: Student of Robotics, Cognition & Intelligence (M.Sc.) at TUM
Julius Hansjakob: Student of Informatics (M.Sc.) at TUM

We also want to credit Daniel Zügner for advising us any time and encouraging to publish this dataset.