Lookout for Vision robustness against variation in training data "good" images

Question

When training a Lookout for Vision model, how feasible is it to include variations in the "good" product training data? As in can I have the product viewed in different orientations (rotated so that a top down view of the product doesn't have to be aligned any certain way), have the product uncentered to the camera, vary lighting?

I am designing a system to identify product anomalies but want to know how much I need to take into account controlling the environment in which the product image is taken when carrying out my design. Ideally I would like the LofV model to be intelligent enough to understand that a product is still a good product even if the top down view of it has been rotated and nudged off centre slightly.

Any help would be much appreciated.

Thanks.

Answer

The Amazon Lookout for Vision [documentation](https://docs.aws.amazon.com/lookout-for-vision/latest/developer-guide/model-create-dataset.html#model-prepare-images) recommends consistent image capture conditions, such as camera positioning, lighting, and object pose. In an ideal scenario you can control the environment and ensure consistent conditions.

If that's not possible, then providing more data that reflects the variations (object position and rotation, lightning conditions) can increase your models' performance. To what extent depends a lot on the situation (variability in conditions, types and sizes of anomalies, etc.) so it's difficult to give a general answer. I would recommend to run a PoC, evaluate your performance metrics (such as [F1 score](https://en.wikipedia.org/wiki/F-score)), and then try to [improve your model](https://docs.aws.amazon.com/lookout-for-vision/latest/developer-guide/improve.html#improve-model).

In addition you can try to preprocess your raw images to reduce variations. For example, you could use an object detection model to locate the object (and its orientation) in the larger image and then crop/rotate the image so it only contains your object (ideally rotated to a consistent orientation).

In your camera setup you could also evaluate bandpass filters as a way to improve lightning conditions. An experienced partner would be a great asset in such a discussion.

Lookout for Vision robustness against variation in training data "good" images

相关内容