This page presents some examples of fluid motion estimates obtained
with our wavelet-based optical flow code. Various configurations are
proposed, featuring particles images or scalar dispersion, small or large
displacements, with various image sizes.
Results are visualized in terms of velocity vectors, or vorticity plot, sometimes superimposed on input images.
Keep in mind that successive image pairs are processed independently, i.e. no dynamic prior is considered at all. As a consequence, overall temporal coherence is a good quality indicator.
Available demos feature PIV sequences:As well as alternative visualizations of 2D turbulence: And synthetic images for testing purposes:
This dataset consists of experimental Particle Image Velocimetry (PIV) pictures of the well-studied case of the cylinder wake at Reynolds=3900. Experiments were realized at Irstea (formerly Cemagref), Rennes, by Anthony Guibert, Johan Carlier and Dominique Heitz. This sequence is made of 3072 frames of 1024x1024 pixels.
Detail of the near wake, sequence of 3 successive pairs of PIV images.
Images have been processed by our optic flow algorithm and are compared to solutions given by correlation-based algoritms such as DaVis or GPiv. The first video below focuses on the near wake, it shows the enhancement brought by our optical flow method (dense fields, bottom row) compared to correlations method (32x32 pixel windows with 50% overlay, upper row). Displayed quantity is the motion vorticity. Second video displays optic flow results for the whole area. Three characteristic instabilities are visible: high frequency Kelvin-Helmholtz, lower frequency von Karman and even lower frequency wake oscillations. Detailled comparison and study of those results is currently in progress.
This dataset consists of experimental PIV images of two concomitant jets, realized in the planar-jet wind tunnel at Irstea (Formerly Cemagref), Rennes (see our experimental facilities) by Anthony Guibert, Johan Carlier and Dominique Heitz. It is a small sequence of 128 frames of size 1024x1024 pixels. The purpose here was to check how large displacements and high shear are handled by the algorithm, and to test the code with relatively large images.
Core region of the flow, sequence of 4 successive pairs of PIV images.
In the following videos, motion is estimated on a Daubechies-5 (i.e. 5 vanishing moments) wavelet basis, along with exact gradient of curl regularization. The first video shows vorticity computed from estimated motion. Second video superimposes velocity vector plot and vorticity contours.
This estimation shows a good agreement with the physics of concomitant jets, featuring the development of Kelving Helmholtz instability in the upper region and a nice tri-dimensionalization of vortices in the lower region.
This is another PIV dataset of 100 frames of size 1024x1024 pixels,
showing a slightly turbulent air flow with small water droplets,
provided by LaVision in the context of the FLUID project.
Core region of the flow, subsequence of 3 successive PIV frames.
In the following video, motion is also estimated on a Daubechies-5 (i.e. 5 vanishing moments) wavelet basis with exact gradient of curl regularization. The first video uses vorticity computed from estimated motion to colorize input PIV images. Second video superimposes velocity vector plot and vorticity contours.
This sequence clearly shows the non-coherence of motion estimated in uniform, badly seeded areas (e.g. the upper-right corner, here). Vortices and structures in the other regions show good spatial and temporal agreement with particle displacements.
This datasets consists in 92 frames of 512x512 pixels, showing the
dispersion of a passive scalar tracer in an electromagnetically-forced
2D tubulent flow. Images were kindly provided by Marie-Caroline Jullien.
This flow being divergence-free like the synthetic one, this estimation uses divergence-free wavelet bases with discrete high-order regularization.
Subsequence of the flow, using 5 successive frames.
At the beginning of the sequence, tracer is concentrated in a small area of the frame, hence resulting in unstable and unreliable estimations in uniform regions (computer visions methods need textured images!). Quality of the estimation increases with the progressive dispersion of the tracer.
High quality mpeg4 (.mov) videos can be downloaded for the vector plot (33 MB).
This schlieren imagery sequence reveals comb-generated turbulence in a quasi-2D flow running on the surface of a soap film. These images were taken at the Laboratorio de Mecánica de Fluidos, Facultad de Ingenieriá UBA, Buenos Aires, Argentina, by the team led by Pr. Guillermo Artana.
Subsequence of the flow, using 3 successive schlieren imagery frames.
The first video below superimposes the curl of estimated motions on input schlieren images. Erroneous motions be noticed that in poorly-textured areas. Second video superimposes a velocity vector plot on curl contours, it emphasizes on the good temporal continuity.
This dataset consists of two synthetic image sequences: particles (PIV)
and scalar dispersion, in a simulated 2D turbulent flow at
Reynolds=3000. Since ground-truth motion is available from the
simulation, error measurement on estimated motion is available. Hence
those sequences are used for evaluation purpose. Each sequence
comprises 100 frames of size 256x256 pixels.
This flow being divergence-free by construction, a divergence-free basis is employed, along with high-order discrete regularization. The video below presents: input synthetic image sequence (left), vorticity of ground truth motion (middle), vorticity of estimated motion (right).