Autonomous Perceptive System Group
Bernoulli Institute for Mathematics, Computer Science and Artificial Intelligence, University of Groningen, The Netherlands
Overview
The use of DJI Phantom 3 Advanced unmanned aerial vehicle (UAV) was used for collecting video frames of cows and
natural backgrounds at different positions and orientations.
The drone was flown three times over different fields containing cows to obtain different samples.
The dataset is called Aerial UAV dataset and was collected by Rik Smit. The dataset contains three cross-splits of unbalanced collections of
two main classes of video frames (cow and non-cow images), with a total of 3981 frames.
The figure below shows some samples of images of the Aerial UAV dataset.
The summary of this dataset subsets and their respective splits are explained in the papers below. Additionally, we provide the raw samples of the video frames and their corresponding meta-files
for creating the image databases (test/train-validation) in .lmdb format. The link to the described dataset can be found here ( download ).
Please note that the training dataset (train-validation)
for each of the cross-splits were further divided into train/validation set.
This condition is only applicable to the deep learning experiments because the trained models were used to evaluate the various test sets.
Rotation Matrix Data Augmentation We propose a new offline data-augmentation algorithm called rotation- matrix data-augmentation (ROT-DA) that transforms an input image to a new single image containing multiple randomly rotated versions put in n × n cells. The use of a larger value for n leads to a new image containing more different poses. The value of n was set to 4 in the experiments, because using higher values of n resulted in making the cow images look very small. An illustration of the proposed data-augmentation method and the overall classification system using the deep learning method is shown in the Figure below.
The link to the described dataset (ROT-DA variant) can be found here ( download ).
Publication
The summary of this dataset subsets and their respective splits are explained in the papers below. Additionally, we provide the raw samples of the video frames and their corresponding meta-files
for creating the image databases (test/train-validation) in .lmdb format. The link to the described dataset can be found here ( download ).
Please note that the training dataset (train-validation)
for each of the cross-splits were further divided into train/validation set.
This condition is only applicable to the deep learning experiments because the trained models were used to evaluate the various test sets.
Rotation Matrix Data Augmentation We propose a new offline data-augmentation algorithm called rotation- matrix data-augmentation (ROT-DA) that transforms an input image to a new single image containing multiple randomly rotated versions put in n × n cells. The use of a larger value for n leads to a new image containing more different poses. The value of n was set to 4 in the experiments, because using higher values of n resulted in making the cow images look very small. An illustration of the proposed data-augmentation method and the overall classification system using the deep learning method is shown in the Figure below.
The link to the described dataset (ROT-DA variant) can be found here ( download ).
Publication
- "Okafor, E., Schomaker, L., & Wiering, M. A. (2018). An analysis of rotation matrix and colour constancy data augmentation in classifying images of animals. Journal of Information and Telecommunication, 2:4, (pp. 465-491)."
- "Okafor, E., Smit, R., Schomaker, L., & Wiering, M. (2017, July). Operational data augmentation in classifying single aerial images of animals. In INnovations in Intelligent SysTems and Applications (INISTA), 2017 IEEE International Conference on (pp. 354-360). IEEE."