Improving
Common problems and how to improve your model.
How can I improve my model?
How can I improve my model?
Your examples are your model’s source of truth about the world. Here are some best practices for building successful projects:
- Understand your problem - start simple, expand over time. Break your scenario down into smaller experiments to prototype and then expand over time.
- Lobe only learns from the examples you import. Try to collect examples that cover the different types of images your model will see and make predictions on in the future.
- More images always help - new and unique images are better. The more unique and different the images are, the better your model will learn to generalize.
- If you can’t classify the label from looking at an image, it will also be difficult for Lobe. Make the image content as large and relevant to your label as possible.
Why is Lobe not predicting well on new images in Use?
Why is Lobe not predicting well on new images in Use?
Compare the images in Use with the images in your examples:
- Your model may have memorized your examples instead of learning to generalize to new images. This is commonly called overfitting. Check Results to see if your model is overfitting.
- These new images are not represented by your examples. Try to play with edge cases and ‘trick’ Lobe as much as possible. Make sure your examples contain all these new variations and conditions seen from Use. You can add new images as examples directly from Use so that Lobe can continually improve.
Why is there always a prediction even when nothing is in the image?
Why is there always a prediction even when nothing is in the image?
Lobe will always predict one of your labels even if your image does not contain any related content. If you expect your model to see these types of images, create a ‘None’ label and add variations of these images as examples. You can use this ‘None’ label as a placeholder when waiting for relevant predictions.
Why are small objects not working?
Why are small objects not working?
The smaller the object is you are trying to classify, the harder it will be for the model to predict. If you can, set up your camera to be closer or zoom in to the object you need to classify. You could alternatively square crop the portion of the image that contains the object you want to classify.
How do I make my model work with multiple objects in my images?
How do I make my model work with multiple objects in my images?
Your model may have a harder time learning with many objects in your images. You should:
- Make your subject as large and centered as possible in your images. The larger the surface area of your image with content related to your label, the better your model will perform.
- Collect more variations of images with multiple objects so that your model can learn to distinguish the objects unrelated to your label.
Why is Lobe predicting one of my labels more often than others?
Why is Lobe predicting one of my labels more often than others?
If you have many more images in one of your labels, Lobe may predict this label more frequently than others. Balance your examples by making sure labels have an equal number of images.
How can I create a new version or keep track of progress?
How can I create a new version or keep track of progress?
Duplicate your project through Project > Duplicate. This will make a new copy of your examples and any training. Experiment with different images or model types on duplicated projects without worrying about losing your original project.
Why are tall or wide images not performing well?
Why are tall or wide images not performing well?
Try to keep your content center frame. Lobe takes the center square crop of every image for training. Important parts of your image may be cropped out and not seen by your model.
How can I make sure my model is not biased towards different groups?
How can I make sure my model is not biased towards different groups?
Examine your examples for potential biases, such as under-representation of minorities or any other images that may have been excluded during collection. Because your examples are the ground truth, Lobe will tend to learn any biases present from your images.
How do I make my model work better under different conditions?
How do I make my model work better under different conditions?
Collect more images in the different conditions that you expect to see in the real world. Your model may not perform well in situations it hasn’t seen in your examples, such as images at night or in different seasons and weather conditions. If you are unable to collect image variations, try to synthetically make them in an image editing tool like Photoshop.