Having large datasets with high-quality annotations is quintessential in any computer vision task involving deep neural networks. Unfortunately, the process of annotating thousands of images is a time and human-resource consuming endeavor. Hence, for many companies and university researchers, the annotation time and scalability become a major pain point to scale their research project or business. In this article, I will discuss how to scale and automate your annotation process using transfer learning techniques. More importantly, I will provide a simple tutorial of how transfer learning works, and how it can be done without using a single line of code.
- What is transfer learning and how it can be applied to the annotation process
- Training new neural networks using SuperAnnotate
- Testing newly trained network
1. What is Transfer Learning (TL) and how can it be applied to the annotation process.
In the most general terms transfer learning (TL), is a direction of machine learning that focuses on storing “knowledge” that a model has learned in order to solve some problem A and use that knowledge to help with another related problem B.
Humans have natural ability to apply knowledge gained from solving one task in order to solve an entirely different one. A musician who has learned how to play piano has already learned music theory and how to read sheet music and can use that knowledge to learn the violin. In other words, if you want to learn the violin you don’t have to re-learn music theory. Transfer learning tries to solve a similar problem in deep learning.
An example that is more related to computer vision is that a neural network that has learned to classify Cats and Dogs in images has perhaps learned useful features that are specific to canines and felines would help another network classify Wolves and Tigers.
Now that we know what transfer learning is, how does it actually help during the annotation process? What we aim at is improving the speed at which image annotations can be done. Let’s look at the typical process of annotating a particular image with bounding boxes and assigning classes to them.
The hypothesis is that if we have a neural network (NN) that can somewhat accurately make predictions on an input image then fine-tuning its predictions and annotating the parts of the image that the NN failed to classify will be much faster than annotating the whole image.
In the example above, I did not try to be very accurate and show that sometimes the annotator will need to readjust and resize bounding boxes. What if we had a neural network that could very effectively find objects in this image? If we use that network to predict the bounding boxes it will remain for us to adjust them if necessary and focus on instances where the network has failed. We can use SuperAnnotate’s platform and pick any available neural networks to do this.
After running the given prediction model the annotated image looks like this:
The improvement in speed will be even more noticeable when we need to annotate the whole semantic mask instead of only bounding boxes.
By leveraging the power the of transfer learning, data augmentation and pre-trained networks we can train new models that solve the task in the domain we are interested with relatively small number of high quality annotated images. Using these newly fine-tuned models to partially annotate images we can tremendously speed up the whole process.
2. Training new NN using SuperAnnotate
At SuperAnnotate there is a possibility to leverage the knowledge learned by well-performing state-of-the-art pre-trained networks in order to create new networks (or to improve the current network) that will suit your annotation needs.
If you are registered in the platform, the workflow for transferring the knowledge from one NN to another would be the following:
- Click on the “Neural Networks” tab.
- Click the “New Model”
- Fill in the model name and model description
- Choose the annotation task and one of the available pre-trained models
- Choose the projects which you want to use for your training (you can choose multiple projects)
- Update some of the default hyper-parameters (optional)
- Choose a GPU to train the new model on
- Click “Run Training”
Among the 6 tasks that are described in the cocodataset.org, we provide pre-trained models for 5 of them`- Instance Segmentation
- Keypoint Detection
- Object Detection
- Semantic Segmentation (will be available in September)
- Panoptic Segmentation (will be available in September)
All available configurable fields for training a new Neural Network.
There are quite a few hyperparameters that can be tuned during the transfer learning process. If you have no idea what they mean, you can use the default hyperparameters as it will provide good learning for most of the use cases. The hyperparameters that we allow to fine-tune are: Batch Size (the number of images used in one iteration of the training procedure), Epoch count, Learning Rate, Gamma (the learning rate gets multiplied by this value after “Epochs for Gamma” epochs), Steps for Gamma, Images (RoIs) per batch (how many regions of interest to suggest per image) Evaluation Period (number of epochs after which a checkpoint of the model is saved, and the performance of the checkpoint is evaluated on the test set) Train Test split ratio (we will use this percentage of images to train the new model)
The user can monitor the training process since we also provide training metrics. Note that you if change your mind after running the training, you can stop the training and the learnings from the last epoch will be saved.
3. Testing newly trained network
Once the new NN model is trained with the given hyperparameters, it can be be used to automate the annotation of the next set of images.
To see qualitatively how the new model performs you need to run “smart prediction” using the newly trained model. A new model with the name that you have specified while running the training will appear in the dropdown list of possible NNs to chose from.
After we create a new model called “New FineTuned Model” using our NN functionality it appears in the dropdown menu for the “smart prediction”
Once the smart prediction has completed you can view the results by clicking on the image and opening the annotation tool.
For one of the clients we have, using a fine-tuned model, observed about 13% accuracy improvement over the original model when trying to annotate instances of “Person” class.
Automating the annotation process without writing a single line of code is essential for many computer vision engineers and annotation service companies. By using SuperAnnotate’s platform, we provided a simple tutorial on how one can set up the automation process using transfer learning, and keep improving the annotation accuracy by annotating and training batches of images over and over again.
Stay tuned … in the upcoming months, we will provide more functionality in our neural network section. More specifically, we will allow you to upload and download your custom models and weights, plot different error metrics, compare and version different training models, etc.