Hey.
I try to detect small objects by using tiling technique.
I noticed that when I do tiling my box loss is increased, and then I noticed there is no overlap parameter available to set when doing tiling.
Is there a workaround, do we set overlap in different step, or it is a feature that RoboFlow still lacks?
1 Like
Hi there @matbyp - the recommended way to do this is with SAHI, which uses overlapping tiles.
Here’s a guide on how to get that set up: Detect Small Objects with Roboflow Workflows
Hey Jacob, I meant tiling during training (pre-process part). when we set a grid for pre-processing with tiling, there is no overlap parameter, which results that training image is sometimes cut through labelled box.
in my particular case my mAP went down because of that. Box loss went down from 0.05 to 0.1, and the improvement in class and object loss did not compensate such decrease.
I’m saying you don’t need to tile during training if you use SAHI during model execution. This solves for your exact cutoff issue, since SAHI intelligently resplices the image.
okay, then what is the benefit of tiling during pre-processing?
with tiling during pre-processing my class and object loss improved drastically. so if we could have overlap parameter during pre-process tiling, the overall performance of my models would go up by quite a bit.
i guess its more of a potential improvements of the current flow.
thanks for replies.
This topic was automatically closed 7 days after the last reply. New replies are no longer allowed.