Evaluation of model performance on the mammalian DeepCell dataset.
<p>(A) Cell-HOTA evaluated at various similarity thresholds α on the mammalian DeepCell dataset, using data from the test set. Five different algorithms are evaluated: Cell-TRACTR, DeLTA, EmbedTrack, Caliban, and Trackastra. (B) Cell-HOTA<sub>0.5</sub> results on the test set. (C)...
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) Cell-HOTA evaluated at various similarity thresholds α on the mammalian DeepCell dataset, using data from the test set. Five different algorithms are evaluated: Cell-TRACTR, DeLTA, EmbedTrack, Caliban, and Trackastra. (B) Cell-HOTA<sub>0.5</sub> results on the test set. (C) Overall results on the test set. Cell-HOTA has three sub-metrics: DetA, AssA, and DivA. OP<sub>CTB</sub> has 2 sub-metrics: SEG and TRA. (D) Comparison of model efficiency. FPS is the inference rate measured in frames per second, so higher values correspond to faster inference times. Params is the number of parameters in each model. M, million. Training time is the number of hours needed to train each model. Since Trackastra uses precomputed segmentation masks (from Caliban), its performance does not reflect segmentation capability, unlike other models, which generate masks as a part of their pipeline. These values are denoted with a * because they only include tracking. Training times are not reported for Caliban and Trackastra because we used pretrained models.</p> |
|---|