Imagga's custom training enables customers to specify the categories to which their content should be assigned and use the Imagga auto-categorization API with the custom-trained model. This is especially useful in the cases where a company has tried to classify content manually but has found it overwhelmingly challenging.
The training process is straightforward.
Provide training data (list of non-overlapping categories and sample image data for each category). In case you do not have sufficient data, our data engineering team will advise you on the available options.
We build a deep learning classification model trained with your specific categories and data. The existing classified content will be used as a training dataset and after completion, the unclassified content will be automatically processed.
The model is then wrapped in an API so it can be easily plugged in and implemented in your company’s existing systems and workflows.
Benefits of Building Custom Models with Imagga.
We have successfully worked with a plethora of companies in various industries, always striving to achieve the best possible results for the image-related business processes of our customers.
Professional Data Collection
In case you don't have the image dataset required for quality machine learning we can help and provide a data collection team.
Robust Computing Power
We have the capacity to handle any size of training dataset thanks to our cutting edge internal infrastructure and computation power.
We can operate the running of the technology in our own super-scalable cloud, as well as assists you in deploying it on your premises or integrate it in your applications on the edge.
Need a custom solution? With our expertise we are helping you to go through the whole proccess.
Training the world’s largest plant recognition model.
There are over 320.000 different species worldwide, meaning that Imagga’s categorizer had to be trained for a huge number of classes. For training, it used over 90 million images, making the scale of this project massive even on an enterprise level.
No, currently this is work performed by our expert machine learning team. We are working on a web interface that will make it possible for you to define categories, upload sample photos and do the training yourself. However our team can be of great help for tasks such as defining the categories' structure, running of the training dataset and the training process for achieving optimal results.
Ideally we need around 1000+ photos per category but 200-500 might work as well if the categories are well distinguishable and each category is well represented by the given photos. The images need to be at least 300px on their shortest side.
The number of photos in each category don't need to be exactly the same, but ideally there won’t be more than x2 times difference between the smallest and largest number of sample photos for their respective categories.
Our training works with flat structures, so it’s best if you flatten the structure. Of course internally you can have your own knowledge of what hierarchical structure your flattened categories belong to.
We are not using your images for any purpose outside of your own project and we don’t share the images in any way. However if you want to contribute to the improvement process of the models, please let us know.
There are several options - adding more diverse sample photos per category can significantly increase the precision rate. Sometimes categories that are overlapping could be the reason for less than optimal results - in certain cases redefining the list of categories would work quite well.
In very rare cases we may jointly conclude that we can’t do anything at this stage and you don’t need to pay the success fee. We try to prevent this case by carefully analyzing the definition of the categorization before we start the actual training and request the upfront fee.