Fig. 5: The role of noise and retraining in generalisation of network performance.
From: Estimating full-field displacement in biological images using deep learning
Displacement vectors from a single frame pair from a video of LECs in a Drosophila pupa evaluated using conventional (a) DIC and (b) OF methods. c Accuracy of methods compared to manual tracking. Error bars are standard error of mean. Displacement vectors from networks trained using different methods: (d) cardiomyocyte data only (excluding Drosophila data); (e) cardiomyocyte data only, with intensity modulation noise; (f) retrained on Drosophila data; and, (g) retrained on Drosophila data with intensity modulation noise. Scale bars are 20 μm.
