Multi-Domain Dataset for Robots (MDDRobots) - Multi-Domain Indoor Dataset for Visual Place Recognition and Anomaly Detection by Mobile Robots

The dataset is password-protected until the article is accepted by the Scientific Data Journal. After acceptance, the dataset will be available to everyone. For any questions, comments or other issues please contact Piotr Woźniak, email: [e-mail].

License

The MDDRobots dataset is made available under the CC BY 4.0 license https://creativecommons.org/licenses/by/4.0/.

Summary

The Multi-Domain Dataset for Robots (MDDRobots) contains data for computer vision problems, indoor visual place recognition, and anomaly detection. The recorded images are from different cameras and indoor environmental conditions. 

It is obligatory to cite the following paper in every work that uses the dataset: 

Wozniak, P.; Krzeszowski T.; Kwolek B.: Multi-Domain Indoor Dataset for Visual Place Recognition and Anomaly Detection by Mobile Robots, Scientific Data, ISSN: 2052-4463, 2025.

Data description

The data are divided into five sets (containing data for different cameras), which have further subsets. Each of the subsets: Training, Test 1, Test 2, and Test 3 consists of nine image sequences. A total of 89,550 three-channel RGB color images in PNG format are organized into 20 zip folders with a whole size of 34.3 GB. Each image in the sequence has a label that represents a room. The number of images for each subset differs due to the split into training and testing data. The difference also results from different methods of recording the image sequences. In order to have balanced data in the subsets, each room in the sequence has the same number of images. Different environmental changes were introduced in each subset. The data from Test 1 are closest to those from the training set. The differences between the sequences are mainly due to changes in the route, robot, and recording equipment. The rooms are well lighted, but not overexposed. The sequences from Test 3 present changed conditions, such as a different time of day, a changed lighting system, and intensive layout changes. The key change is the different paths of the human and the robot. This means a different perspective from previously recorded scenes. The Test 2 sequences pose the most difficult challenge because they contain various recorded activities performed by people moving around rooms. People can occlude important parts of the scene and pass in front of the camera. The images were anonymized by manually blurring the faces of observed people.

Dataset structure

RobotPiCamera_DataSet

DataSet_RobotPiCamera_RGB_train

DataSet_RobotPiCamera_RGB_test1

DataSet_RobotPiCamera_RGB_test2

DataSet_RobotPiCamera_RGB_test3

 Xtion_DataSet

DataSet_XTION_RGB_train

DataSet_XTION_RGB_test1

DataSet_XTION_RGB_test2

DataSet_XTION_RGB_test3

 GOPRO_DataSet

DataSet_GOPRO_RGB_train

DataSet_GOPRO_RGB_test1

DataSet_GOPRO_RGB_test2

DataSet_GOPRO_RGB_test3

iPhone_DataSet

DataSet_IPHONE_RGB_train

DataSet_IPHONE_RGB_test1

DataSet_IPHONE_RGB_test2

DataSet_IPHONE_RGB_test3

P40PRO_DataSet

DataSet_P40PRO_RGB_train

DataSet_P40PRO_RGB_test1

DataSet_P40PRO_RGB_test2

DataSet_P40PRO_RGB_test3

Example folder content: DataSet_P40PRO_RGB_train\Corridor1_RGB - 00000000.png, 00000001.png, 00000002.png, 00000003.png, ... 00000599.png.

Total Images (Images per Place)

Subset
Mounted
Training
Test 1
Test 2
Test 3

Pi Camera
Robot
7200 (800)
5400 (600)
5400 (600)
5400 (600)

Xtion
Robot
7200 (800) 
1800 (200) 
1800 (200)
1800 (200) 

GoPro
Hand
5400 (600)
4500 (500)
4500 (500)
4500 (500)

iPhone
Hand
5400 (600) 
4500 (500)
4500 (500)
4500 (500) 

P40Pro
Hand
5400 (600)
4050 (450)
3150 (350) 
3150 (350) 

Further information

For any questions, comments or other issues please contact Piotr Woźniak .