Uploads
Use the Uploads View in the web platform and or the mobile data collection app to upload images and or videos for the labels that you want to recognize and for the sources under which the images belong.
Required Steps
Collect the minimum amount of required train images for the labels on which you want to train the model - 2 for Few Shot Classifier, 20 for Classic Classifier. Even though 20 is the lower limit, we recommend uploading at least 100 images per label. Also, keep in mind, that for the default source for Classific Classifier, 10% of your data is automatically used as test data, and hence not available for training.
It is recommended to upload a similar number of images for all labels so that you have a balanced dataset for training. However, the training pipeline can handle unbalanced datasets as well.
Notes about File Formats and Size of Uploaded Data:
If you are uploading images smaller than 32x32 pixel, uploads will error.
Supported file formats for images are jpeg, jpg, bmp, png. Supported file formats for videos: mp4, mov, 3gp. Uploaded videos are parsed into images, extracting 1 frame per second.
The image input size for the few shot training pipeline are 224x224 pixels. All images smaller than that, will not be used in training. Larger images will be downsized before training (no impact on performance).
The image input size for the classifier training pipeline are 224x224 pixels, but smaller images are accepted. Larger images will be downsized before training (no impact on performance). Smaller image will be stretched before training - this could potentially have an impact on performance. The more images you have that are smaller than the training input size and the smaller those images are, the larger the impact on performance might be. If you are creating annotations, smaller than the lower limit 20x20 pixels, those annotations will be completely excluded from the dataset and the classifier training pipeline.
If you are using the Passio Data app to upload images or videos, there should not be any issue for none of the use cases, since images taken on mobile phones are usually much larger than 224x224 pixels.
We have a duplicate image check inside of each project. This means, you cannot upload the same image twice in the same project, even if uploading the image with a different label and or source.
Optional Steps
Create a separate test data set under a new source with mode "Test Data for Recognition". Even though the default source for "Train Data for Recognition" has the option enabled to automatically split out 10% of images as test data, it can be helpful to have instead or in addition to that a dedicated test set which exactly depicts your real life use case. The better your test set, the less time you will spend on iterating and testing your SDK.
Collect data for background - collect images or videos where none of the labels that you want the SDK to recognize are included. For some uses cases, it might not be relevant, but for many uses cases it will improve the performance of your classifier model. You can collect background data when choosing the modes "Train Data for Background" or "Test Data for Background" in the upload process. The system will assign the background label for all image data uploaded under sources of these modes.
If at any point you need help from the Passio team, please reach out to us at support@passiolife.com
Last updated