Fig. 2: Performance of globLoc on biplane data.

a Comparison of the z localization error using globLoc and using single-channel fitting followed by CRLB-weighted averaging of positions. GlobLoc improved both the minimum theoretical localization uncertainty (\(\sqrt{{{\mbox{CRLB}}}}\), line) and the localization error (RMSE, circles) by approx. 1.5-fold compared to individual fitting (+, Supplementary Fig. 1). Simulations based on the biplane PSF model from the SMLM challenge 2016, 2500 detected photons in each channel and background of 20 photons per pixel. b Performance of different software on biplane data in the SMLM challenge 2016. GlobLoc (SMAP-global in the submission) outperformed all other software and improved the localization accuracy by 90% compared to the second-best algorithm LEAP²⁵. c Comparison of globLoc under different parameter sharing schemes and compared to ThunderSTORM (detailed settings in Supplementary Fig. 9a) for biplane training data (MT0.N1.LD_BP) of the SMLM challenge 2016. c2-4 Side-view cross-sections of the region as indicated by the dashed white box in c1 (c2: globLoc with x, y, z, photons and background photons shared; c3: globLoc with x, y and z shared; c4: ThunderSTORM.). Red-hot: fitted localizations; cyan: ground truth. d Comparison of globLoc on bi-plane data under different parameter sharing schemes and compared to ThunderSTORM (detailed settings in Supplementary Fig. 9b), on the example of the nuclear pore protein Nup96-AF647. d2-4 Side-view cross-sections of the region as indicated by the dashed white box in d1 (d2: globLoc with x, y, z, photons and background photons linked; d3: globLoc with x, y and z linked; d4: ThunderSTORM.). Representative results are shown from 3 experiments for d. Scale bars 1 µm (c1, c2, d4), 10 µm (d1).