Skeleton accuracy for test data subgroups and types of errors.
<p>(A) Root mean square deviation between model predictions and test data for ‘easy’ cases which were skeletonized by Tierpsy. By construction the number of failed predictions for Tierpsy is zero. (B) The same as A but using manually annotated skeletons as the test data. These data have more d...
محفوظ في:
| المؤلف الرئيسي: | |
|---|---|
| مؤلفون آخرون: | , , , , |
| منشور في: |
2025
|
| الموضوعات: | |
| الوسوم: |
إضافة وسم
لا توجد وسوم, كن أول من يضع وسما على هذه التسجيلة!
|
| الملخص: | <p>(A) Root mean square deviation between model predictions and test data for ‘easy’ cases which were skeletonized by Tierpsy. By construction the number of failed predictions for Tierpsy is zero. (B) The same as A but using manually annotated skeletons as the test data. These data have more difficult cases, but there are still a substantial number that Tierpsy is able to make predictions for (Tierpsy fails around 35% of the time on these data). (C) Examples showing manual annotations (blue) along with DTC predictions in red with different values of RMSD. In all cases, the predicted skeletons are close to the worm midline and the RMSD is driven mostly by overshooting at the head and/or tail. (D) Instead of measuring the point-to-point RMSD, we can measure the distance to the nearest portion of the test segment for each of the model predictions. (E) The mean distance computed as shown in D is more similar across the models although DTC still performs slightly worse.</p> <p>(PNG)</p> |
|---|