Fig 1.
The workflow divides in three mains parts: detection, matching and post-process. The few steps that require user input are indicated by a . Sample dataset: ZFJ_001.
Fig 2.
Two recordings with severe drift are used for benchmarking (top: DRO_001, bottom: ULT_001). (A) Comparison of a frame (magenta) with the first frame (green) and magnification of details in the scene. (B) Root mean square deviation (RMSD) of pixel intensities after registration onto the first image, averaged over all time frames and normalized by the RMSD without registration, for three registration methods. Error bars: standard deviation across time frames. (C) Relative average computation time of the three registration methods, normalized by the total number of pixels in the movie (arbitrary units). Error bars: standard deviation across time frames. Raw data are available in S1 Data.
Fig 3.
Characterization of the TD2 dataset.
(A) Illustration of the dynamics at various timescales in ACT_002. The Voronoï cells (dashed white) and the displacements of a particle at τ = 1, 10 and 100 are overlaid. (B) Geometric probability of incursion pinc (red) and distribution of the reduced displacement ρ at three different timescales τ (black) in ACT_002. The probability of incursion Pinc is the intersection of the areas under the two curves. (C) Pinc as a function of τ for the whole dataset (symbols). The solid lines are fits with a logistic function (see text). (D) Scaling of the reduced quantities Pinc/L as a function of on the standard logistic sigmoid function (solid black). (E) Classification of the movies in the dataset by increasing values of τ1 as defined by Eq (4), with fitting parameters determined over a logarithmic scale for Pinc. Movies with τ1 < 1 are undersampled while movies with τ1 > 1 are oversampled. (F) Comparison of Pinc(τ) for different levels of degradation δ (symbols) and corresponding logistic fits (solid curves) in ACT_002. (G-I) Evolutions of the fitting parameters L, k and τ0 as a function of the degration δ in ACT_002.
Fig 4.
Automatic tracking parameters.
(A) Snapshot and blow-up of ZFJ_001, with definition of and dα (B) Scheme of the algorithm for determining the tracking parameters automatically. (C-E) Distribution of displacements dr (in pixels), angular differences dα (in radians) and area differences dA (in pixels) when the default parameters of the software are used on ZFJ_001, for τ = 1 (black). The corresponding χ and Gaussian fits are displayed in red. Orange bars: resulting soft parameters. (F) Evolution of sr, sα and sA with algorithm iterations for ZFJ_001. Left: iterations 1 and 2; right: iterations 2 and 3. A hundred runs with random initial values are shown, the run with the software default parameters is highlighted in red. (G) Evolution of Pswap with algorithm iterations, same runs. (H-J) Evolution of the converged parameters
,
and
as a function of the timescale τ for ZFJ_001. (K) Comparison between Pswap (blue crosses) obtained with the converged parameters and Pinc (red dots) for ZFJ_001. The solid black line is the logistic fit of Pinc.
Fig 5.
Benchmark of FastTrack, idtracker.ai and ToxTrac.
(A-B) Comparison of the computation time for the tracking of various movies with the same workstation. Whenever possible, CPU and GPU variants of idtracker.ai have been run. Only the first 100 images of DRO_002 have been used. (C-D) Accuracies of the resulting trackings. “perfect” means an accuracy of exactly 1. The trajectories computed by the CPU and GPU variants of idtracker.ai being rigourously similar, we only show the results for the GPU. For 100Zebra, the accuracy was computed by taking into account only the first 200 images.