How important is stride size when exporting data for an Object Classification Deep Learning model?

[ScarletMaguire]

I am creating some training samples for an Object Classification deep learning model, which I will then export. I know that chip/tile size is important because they need to be big enough to cover the objects of interest. However, how important is stride size? I know that for Object Detection, stride size is important because it ensures that there is some context surrounding the objects to be detected. However, for my model, the objects do not need to be detected as they are all already delineated, they only need to be classified. Also, for my model in question, the surroundings of the objects are not relevant to their classification.

So, when exporting the data, should I bother with a stride size or just leave this parameter blank?

[ShivaniPathak]

Hi @ScarletMaguire , for your object classification model, the stride size is generally less critical compared to an object detection task. In object classification, you are focusing on the classification of already delineated objects within the tiles, and the surrounding context typically isn’t relevant.

• Stride size: In classification, this defines how much overlap there is between adjacent tiles when you're extracting patches. If your objects are well-contained within the tiles and there's no need for context around them, the stride size becomes less important.
• When to adjust stride size: If your objects are small and close to the edge of the tile, a small stride could be helpful to avoid losing important features at the tile boundaries. However, if you’re confident the objects are completely contained within each tile (with no overlap), you could leave the stride size blank or set it to match the tile size (no overlap).
• In short: For your use case, the stride size can often be left at default (or blank), and it won't negatively impact the model. However, if you think the objects may be near tile boundaries, a small stride size (or overlap) could be beneficial.

This should provide enough flexibility for your classification task without requiring much focus on stride size, unless you encounter specific edge cases with object boundaries.