Deep learning is a sub-category of machine learning methods based on artificial neural networks that automatically extracts high-level features from the raw input data. Between different types of deep learning methods, convolutional neural networks (CNNs) are among the most successful and commonly used architectures in the deep learning applications. Therefore, here in saiwa two CNNs are provided for training on user specific dataset (i.e. Detectron2 and Yolov5). Please check the application of Detectron2 and YOLOv5 on object detection and count objects services. For more information about deep learning concepts, please refer to the corresponding white paper.
There are two ways of providing the training data (raw input plus annotations): 1. Directly uploading from user computer or cloud space and 2. Setting a link to a public dataset. In the second method saiwa© team will download the dataset. For annotating the images, saiwa boundary (for Detectron2) or bounding box (for YOLOv5) annotation services is useful. We can also annotate the images for you. You only require to select “not annotated” option in demo form.
After setting the job, you may follow up the training process and see the results from your panel. When training is done, the model is ready to download from user panel and also is accessible from object detection and count objects service interfaces. For any problem in setting job or tracking the train process or testing the trained model, please leave us a ticket.
saiwa deep learning service provides the following features and facilities:
- Automatically train on user specific data
- Validating the trained model from other services like object detection
- Train on public datasets with minimal user involvement
- Compatible with saiwa boundary and bounding box annotation services
- Request for annotation by saiwa© team
- Tracking training jobs from user panel with diagrams and other statistics
- Editing the number of epochs after the model reached use desired validation accuracy
- Download and online test of the trained model
- Compatible with saiwa object detection and count objects service for test the trained modelThe two networks that are currently available for training on user specific data (i.e. Detectron2 and YOLOv5) are applicable for detection and segmentation problems, like:
1. Object detection
2. Object localization and segmentation
3. Object tracking
For more specific applications, please refer to object detection and object counting service descriptions.