to intelligently merge segmentation predictions from two different sources, I've implemented an ensemble method that prioritizes geometric quality. Instead of relying on arbitrary averaging or voting, this approach evaluates the shape of each predicted polygon using the isoperimetric quotient—a smoothness metric calculated from the perimeter-squared-to-area ratio.
For any given object, the system compares the two predictions and automatically selects the one with the smoother, more regular geometry, which serves as a proxy for a higher-confidence prediction. first model IOU is 0.957 Second 0.956 and ensemble is 0.958
'/%3e%3c/svg%3e)
Same here! Makes you wonder how those top scores are being obtained, can't wait to see their approach, some people are just extraordinary 🔥
I think it comes from a single model if you ask me. Have you tried weighted ensembling?
I use smooth polygon detection ensembling and it works for me
to intelligently merge segmentation predictions from two different sources, I've implemented an ensemble method that prioritizes geometric quality. Instead of relying on arbitrary averaging or voting, this approach evaluates the shape of each predicted polygon using the isoperimetric quotient—a smoothness metric calculated from the perimeter-squared-to-area ratio.
For any given object, the system compares the two predictions and automatically selects the one with the smoother, more regular geometry, which serves as a proxy for a higher-confidence prediction. first model IOU is 0.957 Second 0.956 and ensemble is 0.958
Impressive. Unfortunately I don't think there is enough time for me to try something like that.