Datasets

Data, the fuel of proper model training

Data have been called the new oil of the digital economy; with the rise in AI applications in almost every aspect of human life, nothing else is expected. The word data is used to refer to many things, but in applications of computer vision, the word broadly embodies the collection of images that is used to train a model.

Where to start when researching on cattle?

Whether you want to create a novel architecture for cattle analysis or if you just want to apply a currently-established method to your work, you would face the challenge of data acquisition. Finding the best data that can fit your needs can take a considerable amount of time and even more when you realize that the data requires preprocessing and labeling.

Given these challenges, it is always a good idea to build upon the achievements of previous research. By looking through the publicly available datasets and models, you may be able to fast-forward this process and facilitate your development process.

The Hub aims to make this process easy. We have gathered a list of publicly available datasets with description for their applications. We will soon provide a uniform model haven that allows you to seamlessly download and work with these datasets. You can find the information about these datasets in the table below.

Our team is also on the lookout for any new research that introduces new and exciting datasets and will be adding them to the list.

What to do if available datasets are not enough?

One of the main goals of the Cattle Open Hub is to highlight gaps and scenarios where there is not enough data available. We have already realized that datasets from crowded environments, such as feedlots, are lacking and we have taken initiative to solve this issue; as a milestone of the Hub, we aim to publish our dataset on crowded pose estimation in cattle (you can find samples below). The main goal of the Hub is to connect academia and industry; So with identified gaps, ranchers can easily provide researchers with the required data. Moreover, by testing the models in real-life, ranchers can provide their feedback and expectations directly to a group of researchers, and make sure that the products are going to be addressing the challenges they have.

Our Pose Dataset

We have developed a custom pose estimation dataset by synthetically generating environments and placing cattle within them. For more detailed information, please refer to our publications. You can request access to the dataset here: https://forms.gle/qjHXNF8ejoRf3DrV7

Synthetic Data, a Workaround Solution

Another promising avenue of research is the generation of synthetic data and data augmentation. With synthetic data, researchers can expand their datasets, allowing them to train larger models. Synthetic data can also be used to create instances of data from scenarios where data collection is costly, time-consuming, or is not possible on a large scale. Synthetic data generation in the ﬁeld of animal pose estimation has seen some advancements in the past couple of years, and cattle pose estimation can beneﬁt from it.

Page updated

Google Sites

Report abuse