Pseudo-real Image Sequence Generator for Optical Flow Computations

The availability of ground-truth flow field is crucial for quantitative evaluation of any optical flow computation method. The fidelity of test data is also important when artificially generated. Therefore, we generated sequences of artificial flow fields together with an artificial images based on given real-world sample image. The presented framework benefits of a two-layered approach in which user-selected foreground was locally moved and inserted into an generated background based on real image. The background is similar to input sample image while the foreground is extracted from original and so is the same. The framework is capable of generating 2D and 3D image sequences of arbitrary length. Several examples of the version tuned to simulate real fluorescent microscope images are presented. Example of a verification of an optical flow computation method. Computed flow field between given two frames is visualized in red vectors whereas ground-truth flow field is shown in green vectors. Example of a change in computed flow fields when computed with the same method and the same parameters on changed image data. In the right column from top: the sample image, the sample image smoothed and the sample image with noise added. In all three cases, the scene is still recognizeable and the motion is still the same but the performace of investigated method differs. → → → → → → MOTIVATION and TERMS The optical flow is, according to Horn and Schunck [1], the distribution of apparent velocities of movement of brightness patterns in an image. In other words, the outcome of an optical flow computation method, applied on two consecutive images, is a flow field in which a velocity vector is assigned to every voxel in the image. The vector represents movement of a given voxel from the first frame to the second. This is often used for representation of a movement in the sequence of images [2]. In particular, optical flow methods are often used for analysis of time-lapse image sequences acquired from fluorescence optical microscope [3,4]. The most common approach to validation of a method for computing optical flow is to compare its result to some certain flow field [5] that is commonly termed as ground-truth flow field. We don’t have the ground-truth information at hand when testing on real data, unfortunately. Therefore, automatically generated pseudo-real high-fidelity test images together with correct flow fields may prove very useful. For example, vast amount of unbiased data may speed up the tuning of an existing or newly developed optical flow computation method by allowing for its immediate evaluation over close-to-real data. Since we considered 3D image as a stack of 2D images, we didn’t utilize any 3D-to-2D projection — flow field remained truly 3D. Thus, we refer to such a 2D or 3D flow field as to a ground-truth flow field. The three images illustrate idea of optical flow. The left-most and the right-most images are related through the flow field in the middle. The flow field describes the movement of each pixel from the first (in the left) to the second (in the right) frame. If movements of all pixels from the first frame were described in advance, we could compute the second frame by forward transformation [6]. Or, we could compute the first frame by backward transformation [6] if the second frame were given — like we did in this illustration. The flow field describes the movement given by superposition of rotation 3deg clockwise around a cell’s centre, translation by (2px,-2px) and four local translations of foreground regions with distances up to 3px. Anyway, notice artifacts in the left image. Colour coding of flow field representations: the colour determines the direction of given flow vector, the intensity determines the length of given vector. I H G F