Like many developers and engineers in the field, Isabel started her journey in computer vision in academia. For some years she was working on image processing for biomedical images in the pre-Deep-Learning-era using handcrafted features. About three years ago she was offered a position at a start-up to apply deep learning on real-life projects. Isabel thought it should be quite similar considering that she was working in applied research before… well she was wrong.
Discover 3 main lessons she learnt.
The three main lessons Isabel learnt:
- Collect all relevant technical details. Because of our excitement and the hype it’s easy to get ahead of yourself but it’s absolutely necessary to know all technical details in order to find the right model.
- Choosing architectures is quite different since many of the measures for models do not apply in such cases. Many customers are not interested in mean Average Precision and also GPU inference might not be applicable in cases such as running models on smartphones. Another problem is that research benchmark datasets like COCO or Imagenet tend to be quite far from real life datasets. Other differences can be also observed in input data or bounding box sizes for object detection. In the end, performance plots can look vastly different compared to plots found in research papers.
- Testing on a fresh batch of a large size is crucial for finding out if the chosen model works in production / deployment. Academic standards such as splitting the set to 75% of the data for training and 25% for testing might be sufficient for model validation but is generally insufficient for forecasting how good the model will perform in the wild.