Fig. 6.

Example fin candidates and predicted quality ($F_{\mathrm{inst}}^p$). (Top) candidates and their scores after non-maximum suppression. (Other) Candidates and scores from region around the fin before non-maximum suppression. The predictive ability of the model is reflected in the stroke quality predictions for strokes describing at least part of the fin. It is unsurprising that the model makes high quality-predictions for the caudal fin stroke. We also see that while higher scores are sometimes predicted for purely background objects, the scores predicted for these are typically not as high as those predicted for good quality strokes describing fins themselves