When running computer vision models on mobile devices, the focus switches from raw performance metrics to user experience. On mobile phones, this means minimizing battery and disk usage: we don't want to drain the phone's battery in minutes or fill up all the available space on the device. When running on mobile, it is recommended to use smaller models. As they contain fewer parameters, they use less disk space. Moreover, as they require fewer operations, this leads to reduced battery usage.
Another particularity of mobile phones is orientation. In training datasets, most pictures are provided with the correct orientation. While we sometimes change this orientation during data augmentation, the images are rarely upside down or completely sideways. However, there are many ways to hold a mobile phone. For this reason, we must monitor the device's orientation to make sure that we are feeding the model with images that are correctly...
