FAQ

Who can use the Mobile AI platform? Do I need any previous knowledge?

For the last step, the SDK integration, mobile engineering experience on either iOS or Android is required. The rest of the steps can be completed without any coding experience. For step 1, the domain architecture and step 4, the training, ML, AI or data science knowledge are not required, but will certainly be helpful to understand the process more quickly and iterate more quickly. In general, the platform is designed for users without prior experience in ML.

Do I have to go through all data and mark it as reviewed? Or is it okay if we can keep the state of images as unreviewed before training our model?

You do not have to go through your data and mark it as reviewed, the curation step is optional. In the dataset creation process, you have the ability to check a box whether to include unreviewed or only accepted annotations.

Do we have to have balanced data before training our model?

You don’t have to necessarily have a balanced dataset before training - the training pipeline can handle unbalanced datasets. However, the more imbalanced your dataset is, the larger could be the potential impact on the model performance.

How to access the label unknown using the data collection app?

The label "unknown" is a special label being available by default in every project. You cannot upload anything actively for that label in the data collection app because you should attach a specific label to everything you want to upload. Unknown is just a fallback label in case something goes wrong during the upload process. You can of course create a new label called “mixed” or “unclear”, and use that label for uploading image data.

Doesn’t Image classification require a lot of data? How much time will I need?

For the Few Shot Classifier, there is very little data required (2-10 images per label). For the Classic Classifier, you’ll need to collect data. How much depends on the number of labels you want to recognize and the difficulty of recognizing them. If you have a use case of 6 different labels and an average of 100 images per label, it might take you only few hours to complete all the steps, from label definition to packaging the SDK. The minimum time it will take to train your model is on average 20 minutes. The more data you have, the longer data collection might take and the longer the training time will be. The most important factor in collecting data efficiently is knowing what data you want to collect and where to find it.

What kind of data is good to collect?

You can collect your own data using the mobile companion app, available for iOS and Android, or upload existing image data via the web platform. For the Few Shot Classifier, very typical representations of your labels are recommended (2-10 images required). For the Classic Classifier, the Mobile AI platform will prompt you to collect at least 20 images per label, but we recommend at least 100 images per label for the best results. In general, the more data, the better your results will be. The data you collect should also reflect what you expect users to see when they use your app. A variety of orientations, angles, and lighting conditions will ensure that your users get good results.

What would the Mobile AI SDK be good at? What are some example uses cases?

The SDK you’ll make will be good at classifying between a few different options using visual characteristics, like shape and color. E.g. some classification apps in the app store today can identify trees by their leaves or insects. All classification happens on-device, so your app won’t need wifi or cell data. This would be great for cases when the user is away from their home or office.

How many people do I need on the team to build an SDK?

You can be a one person team and complete all the steps to create the SDK. For using the SDK inside of an application, some mobile engineering experience is required. There is no limit to the amount of team members you can invite to your platform organization. You can invite additional team members, if you want to include more people into the data collection process, or just invite them to view the resources and progress. Step 2 - collection - is the step that is most suitable to split between a larger team. You can also let separate team members work on separate projects at the same time. However, keep in mind that we are not yet surfacing any dashboard statistics on team members. This will be included in future versions.

Can I train several datasets at the same time?

Yes, you can run several trainings at the same time.

Can I delete a project?

Yes, you can delete any project if your account has the appropriate permissions. Deleting a project will delete all of its content and you will not be able to recover any of it. Use the Delete button in the Home view, which is at the right end of each listed project.

Can I delete my organization?

Only the owner of the organization can delete the organization. Deleting an organization will delete all of its content and you will not be able to recover any of it.

Do you have any suggestions about the collection of background label? What data is good for background data? Depending on your use case, different kind of objects / contexts might be more relevant to include into background than others. In the case of food recognition, it would be important to upload images of hands or kitchen areas to the background label, because those will be usually present when people scan food, and you don’t want to recognize those as food. You could mimic behavior of people using your application and record videos for your background data, that don’t contain any of the labels you want to recognize, but contain scenery or objects that are likely to be present in the scanning process. Also, you could search for CC0 datasets (e.g. on kaggle) and upload those to your project.

Does the platform offer any methods to preprocess the images that go into training? Like modifying the images to grayscale?

There is no preprocessing done before or during upload to the platform. However, when you create a dataset and train on that dataset, the training pipeline will do some processing and augmentation depending on what is best fitted for the training pipeline. Those augmentations/processing are not visible for you on the platform. We do not yet offer the option to users to modify or customize the training pipeline, e.g. to choose a training pipeline where specific grayscale processing is applied.

Does the platform itself include any sort of images, e.g. for the background class?

The platform currently does not provide you (yet) with any image data, that you could use for your labels or background. You don’t have to necessarily include background into the training, for some uses cases it might not be essential to have background, but for most uses cases it will improve the performance.

Why are some of the images in the Browse Annotation view slightly distorted?

The annotations you can see here can be slightly distorted, because they are being resized to a square, for model training reasons. If you click on the bounding box icon at the bottom of the image, you can see the full size image, and also where the bounding box position of the annotation is set.

I have issues with using my Apple ID for signing into the mobile app on my iPhone

Make sure when setting up the sign in to the mobile app with your Apple ID on your iPhone, you have selected the checkbox: Share My Email, and not Hide My Email. If you already have selected Hide My Email, go to Settings>Click on your Apple ID Profile>Password & Security>Apps Using Apple ID. Here you see a list of all apps that use your Apple ID for sign in. Select the Passio app (bundle ID: com.passiolife.universalCollection) and select "Stop using Apple ID". Now, reopen the Passio mobile app and select Sign in with Apple and make sure to select Share My Email.

I tried to sign into the mobile app with a different email and now it won't let me select the correct email to sign in and I see "user not found" instead.

Try to clear the cache of your (default) web browser on your phone. E.g. if you are using Chrome: Open the chrome app>Settings>Privacy>Clear Browsing Data. If open, close the Passio Data App. The next time you will open the app and login, you should have the email choose menu available again.

Who can see the images I upload to the platform? What solutions do you offer to the users to protect their privacy online through your tool?

All the data that you upload to Mobile AI is private and Passio follows contemporary data security best practices. Your data is accessible only to you and the people you invite to your Organization.

Is it possible to invite more team members into a project? Or is a project only accessible by a single user on the platform?

You can invite additional users in the "My Team" view, under the profile pulldown in the upper right.

How to accept a large number of images?

If you want to bulk accept images, you can do so from the Annotations > Browse Annotations or > Curator view. Hold shift to select a range of images and then click the Edit button in the button bar above the gallery. The dialog that comes up will tell you the number of images you've selected and you'll be able to set their state to accepted. However, you don't necessarily need to accept images - the curation step is optional. You can train, by default, on both accepted and unreviewed images.