Dense Disparity Estimation in Ego-motion Reduced Search Space

Transcription

1 Dense Disparity Estimation in Ego-motion Reuce Search Space Luka Fućek, Ivan Marković, Igor Cvišić, Ivan Petrović University of Zagreb, Faculty of Electrical Engineering an Computing, Croatia ( Abstract: Depth estimation from stereo images remains a challenge even though stuie for ecaes. The KITTI benchmark shows that the state-of-the-art solutions offer accurate epth estimation, but are still computationally complex an often require a GPU or FPGA implementation. In this paper we aim at increasing the accuracy of epth map estimation an reucing the computational complexity by using information from previous frames. We propose to transform the isparity map of the previous frame into the current frame, relying on the estimate ego-motion, an use this map as the preiction for the Kalman filter in the isparity space. Then, we upate the preicte isparity map using the newly matche one. This way we reuce isparity search space an flickering between consecutive frames, thus increasing the computational efficiency of the algorithm. In the en, we valiate the propose approach on real-worl ata from the KITTI benchmark suite an show that the propose algorithm yiels more accurate results, while at the same time reucing the isparity search space. Keywors: stereo vision, visual oometry, isparity estimation. 1. INTRODUCTION Depth estimation plays an important role in many autonomous systems in automotive inustry, augmente reality, an robotics in general. Besies time-of-flight cameras, which can be limite by external lighting an range, stereo cameras are often use as a primary vision sensor for epth estimation. Even though using cameras requires significant computing efforts to obtain isparity maps (DM), stereo cameras are a very popular choice ue to their commercial availability an efficiency. In orer to obtain valuable epth information, images from calibrate an synchronize stereo cameras are use. Early research was focuse on local methos that consiere only neighboring pixels to fin a stereo match. Lack of rich texture in the stereo images resulte with semiense DM with poor accuracy. To tackle these problems, pixel-wise global methos penalizing iscontinuities were introuce, where a epth map is sought minimizing a global energy function, e.g., works of Kolmogorov an Zabih (2001); Klaus et al. (2006); Yang et al. (2009). However, global methos incur high computational an memory costs an in Hirschmüller (2008) semi-global matching (SGM) algorithm was propose. Therein the computational complexity of global optimization methos was reuce by reucing the omain of consiere pixels to several linear paths in the images. Even though the This work has been supporte by the European Union s Horizon 2020 research an innovation programme uner grant agreement No an by the Croatian Science Founation uner contract No. I This research has also been carrie out within the activities of the Centre of Research Excellence for Data Science an Cooperative Systems supporte by the Ministry of Science, Eucation an Sports of the Republic of Croatia. DM obtaine by SGM is ense an more accurate than the ones obtaine with local matching techniques, they still lack temporal information an can introuce epth flickering among neighboring frames. As oppose to stanar approaches that extract epth information from a single stereo image, Dobias an Sara (2011) preict the DM using ego-motion an DM of the previous frame. They take preicte isparities as they are an fill the rest of the DM using a traitional stereo algorithm, while checking the valiity of preicte isparities without any metho to refine them. Aitionally, Jiang et al. (2014) etect moving objects in the scene an avoi preictions base on ego-motion in these regions of the image. However, their algorithm accumulates propagate isparity error with time, since they also take preicte isparities without any refinement. To aress this problem, the authors iscar the preicte DM every 100 frames to start from a fresh SGM-base DM. From the viewpoint of accuracy, theoretically this result oes not outperform SGM, since unreasonable isparities are propagate into future frames until the reset is performe. From estimation perspective, Vaurey et al. (2008) integrate previous frames with the current one using an iconic (pixel-wise) Kalman filter as introuce by Matthies et al. (1989). Their work extens the iea of integrating stereo iconically to provie more information with a higher certainty. They exten the Kalman filter moel by introucing isparity rate in the epth irection. However, they constrain their moel to motion in longituinal irection an, thereby, neglect all movements in lateral an vertical irections. Their algorithm performs well only in scenes where most of the movement is in the longituinal irection, such as highway traffic scenes. Morales an Klette (2013) aim

2 to improve the isparity estimation of objects that are static with respect to the groun or moving longituinally away from the ego-vehicle. However, they also constrain their motion moel (cf. Franke et al. (2005)). Jakob Engel, Jörg Stückler (2015) consier ego-motion an form keyframes that contain epth information for simultaneous localization an mapping (SLAM). They integrate the isparity information from current frame into the assigne keyframe. Since the focus of their work was SLAM, DM generate by their approach are semiense. Žbontar an LeCun (2015) use convolutional neural networks (CNN) to learn a similarity measure on small image patches an compute matching costs. SGM is then use to optimize the results. Mayer et al. (2016) formulate the problem as a supervise learning task that can be solve with CNN. They propose three synthetic stereo vieo atasets for training of large networks an presente a CNN for real-time isparity estimation that provies state-of-theart results using high power graphics processing units. In this paper we propose to generalize the isparity preiction moel an increase the accuracy of epth estimation in comparison to stanar methos like the SGM. We focus our work on stable, precise an fast spatio-temporal reconstruction, thus constraining the usecase of the propose metho to static scenes. Although this can be seen as a limitation, in fact, this approach will form the base for ense stereo etection of ynamic objects an can also be use in applications where static scenes are preominant, e.g., space exploration robotics. We track pixels from the previous stereo frame to the current one irectly in the isparity omain using ego-motion estimation. In orer to avoi the nee to introuce any other sensors, we obtain ego-motion using the visual oometry algorithm propose by Cvišić an Petrović (2015). For each pixel we eterministically compute the isplacement base on egomotion an stochastically track the value of its isparity while upating its uncertainty through time with Kalman filtering. Disparity of each pixel is estimate by combining the newly matche (measure) DM an preicte DM. We perform stereo matching using a custom SGM on a reuce isparity search space, base on the preicte DM an its uncertainty to reuce the computing complexity an number of matche outliers, while proucing enser an more accurate DM. In the en, we valiate the propose approach on real-worl ata from the KITTI ataset (Menze an Geiger (2015)) 2. ALGORITHM DESCRIPTION As an input for stereo isparity estimation, we use a sequence of stereo image pairs, capture using a pair of calibrate, rectifie, an synchronize cameras. The propose algorithm relies on using the previous an current stereo image, I k 1 an I k, with the accompanying DM of the previous frame, D k 1, in orer to estimate the DM of the current frame D k. Each stereo frame at time instant k consists of a left an right image, IL k an Ik R, an the accompanying left an right DM, DL k an Dk R. We use two consecutive stereo images, I k 1 an I k, for ego-motion estimation of the stereo rig. Given the estimate isplacement calculate from I k 1 an I k, an the DM D k 1 from the previous frame, we can preict the DM of current frame D k k 1. This isparity preiction serves as the base for a stereo matching technique, e.g., in the present paper we use SGM, to reuce the isparity search space, thus consequently reucing the computing efforts an the number of outliers in the measure (matche) DM D k m. To increase the accuracy an prouce a enser DM, the preicte DM, D k k 1, is upate with the measure DM, D k m, within the framework of the Kalman filter. This step prouces the upate isparity D k k an we enote it as the isparity upate. The algorithm can be summe up as follows: (1) Ego-motion estimation (2) Disparity preiction (3) Stereo matching (4) Disparity upate. 2.1 Ego-motion estimation In orer to preict current DM D k k 1 base on previous frame DM D k 1, an ego-motion estimation is neee. Transformation Tk 1,L k represents the homogeneous transformation from the current coorinate frame of the left camera, FL k k 1, to the previous frame FL. This transformation can be obtaine using several ifferent methos. In the present paper, we choose the visual oometry approach. Disparity space. We perform preiction irectly in isparity space an, therefore, we transform Tk 1,L k from Eucliean space to isparity space. Let w, M an Γ be efine as x X f y Y 0 f 0 0 ω =, M = Z, Γ = fb, (1) where x, y an are coorinates in the isparity space, X, Y an Z are coorinates in the Eucliean space camera frame FL k, b is the baseline of stereo rig an f is the focal length of the cameras. Γ represents a projective transformation between homogeneous coorinates M in Eucliean space an homogeneous coorinates ω in isparity space (Demirjian an Darrell (2001)). Analogous to the stanar homogeneous transformation in Eucliean space M k = Tk 1M k k 1 (2) the transformation of coorinates in isparity space is efine as ω k Hk 1ω k k 1 (3) where enotes equality up to a scale factor an H is efine as Hk 1 k = ΓTk 1Γ k 1. (4) 2.2 Disparity preiction Using the ego-motion Hk 1 k we transform each pixel from D k 1 to estimate D k k 1. The transformation is performe irectly in the isparity space using (3). Note that in (3) the result of each pixel transformation is only equal up to a scale factor. Since ω is expecte to be homogeneous as in (1), we scale ω k by iviing it with the value of its fourth member to get a homogeneous coorinate form ω k.

3 Uncertainty preiction. While new x, y an coorinates of each pixel are eterministically calculate by (3), each isparity k 1 in D k 1 is associate with an accompanying variance p k 1, forming a variance map P k 1 that inclues left an right DM variances P k 1 L an P k 1 R. Variance of each isparity is preicte using the motion moel applying the same isplacement = (Φ k 1 ) 2 p k 1 + q k 1, (5) where q k 1 enotes the variance of estimate ego-motion p k k 1 in the isparity space, an Φ k 1 enotes the system moel of current transformation for a pertaining pixel. The variance of estimate ego-motion epens on the precision of the use oometry algorithm. How to compute oometry variance an transform it to isparity space is out of the scope of the present paper, an in lieu of a time varying q k 1, we use an empirically etermine constant q. Since we alreay know k k 1, for variance estimation we compute the system moel as the following ratio Φ k 1 = k k 1 k 1 (6) thus avoi the nee for computing Φ k 1 analytically. Disparity refinement. When transforming each pixel from D k 1, it is likely that multiple pixels from frame k 1 will result in the same x an y coorinates in D k k 1. If there are no outliers in the preiction process, pixels with the highest isparity value k k 1 are the closest ones to the camera an are most likely the pixels not being occlue by other pixels. Therefore, we select these pixels an iscar the others. Since we take a eterministic approach to x an y coorinates propagation escribe with (3), incorrect isparity preictions are expecte near object eges. Using egomotion information T k k 1 an previous isparity k 1 we preict the observation of k k 1. In case of ba ego-motion estimation, preiction k k 1 coul result with wrong isparity value. If the isparity is place away from epth iscontinuities, isparity preiction error will be small enough an stereo matching in a close interval aroun this isparity will easily correct the wrong preiction (see Section 2.3). On the other han, if we consier isparity k 1 that is place on the epth iscontinuity (ege), preiction base on ba ego-motion coul result with isparity error that can not be correcte in stereo matching phase. To aress this problem, we reject all the preicte pixels near isparity iscontinuities. Aitionally, if the stereo rig is moving forwar, gaps in the preicte DM appear. This phenomenon can be compare to the zooming effect when lack of information results in holes in the preicte isparity (confer Fig. 3b for an illustration of this effect). We aress this problem with task-specific interpolation that fills the invali isparities cause by this phenomenon. We use horizontally neighboring pixels an use their values to etermine the value of an invali isparity. If i is an invali preicte isparity in an image row, its value is etermine as i = i 1 + i+1, if f s ( i 1, i+1 ) = 1 2 (7) invali, else where f s represents a similarity function escribe as Fig. 1: Variance estimation by counting neighboring isparities from left an right sie of the cost function aroun its minimum. { 1, if a b < γf f s (a, b) = (8) 0, else where γ f enotes a threshol value efining if a an b (in our case, i 1 an i+1 ) are similar enough. In other wors, we compare neighboring isparities an, if they are similar enough, we use their values to fill the invali pixels. We take the same approach for filling invali values base on vertical neighboring pixels. 2.3 Stereo matching Once the preicte DM, D k k 1, an its variance map, P k k 1, is available, we use this information to aitionally reuce the computational complexity of the process of stereo matching. For the stereo matching we use SGM with eight-path configuration. For each path an energy function that penalizes isparity changes among neighboring pixels is minimize to fin the optimal isparity value for each pixel. Unlike its usual implementation, we reuce the isparity search space base on the preicte DM D k k 1 an P k k 1. Instea of searching the whole isparity space for every pixel at every iteration of stereo matching, we only consier small environment centere aroun the preicte isparity k k 1. The reuce search interval is efine as [ k k 1 ± 3 p k k 1 ]. Matching uncertainty. To estimate the variance of each matche pixel, we refer to the approach escribe in Weel et al. (2011) where authors observe the slope from left an right sie of the cost function aroun its minimum. The slope serves as a quality measure of the estimate isparity. If the slope is low, the isparity is not estimate precisely an the variance is higher. If the slope is high, the isparity is estimate precisely an the variance is lower. Weel et al. (2011) have shown that uncertainty etermine using this technique is correlate to the true variance of the errors base on comparison with groun truth. The isavantage of this approach lies in the case when there are two or more neighboring pixels with minimal cost. This way the slope is horizontal an the uncertainty is infinite which often is not the case. We propose to expan the variance estimation approach by counting neighboring isparities n l an n r while sums S l an S r of their costs are lower than S max (Fig. 1). In our experiments, we etermine that for matching winow [3x3], S max = 10 performs well over a wie range of scenes. The matching variance r k is then efine as

4 r k = (n l + n r ). (9) By consiering multiple nearby matching cost, we avoi infinite uncertainties. Aitionally, the variance value oes not epen on the immeiately neighboring costs alone, but also on the costs that fit into sum S max. Alternatively, innovation variance s k = p k k 1 + r k coul be use for isparity search space reuction. However, in orer to etermine s k, we woul nee to first calculate the matching variance r k. Since r k epens on the isparity value an nearby optimization function costs, which are available only after the matching process, we woul nee to rerun the matching process (on a wier interval), which woul inuce aitional computational loa. Temporal stability. In Jiang et al. (2014), isparities are propagate from previous moment k 1 to current moment k as they are. If an unreasonable value (outlier) is present in the previous frame, it will also be propagate to the current frame. Since there is no mechanism to forget these unreasonable values, they will be propagate to every future frame an the number of outliers in the final DM will increase with time. Jiang et al. (2014) aress this by forgetting the preicte isparity every 100 frames. In this work, we aim at continuous an constant improvement in accuracy an computing time, hence no preicte pixel is taken as it is. Preicte value is only use to reuce the isparity search space an its variance is tracke for further filtering as escribe in Section Disparity upate Combining the preicte DM D k k 1 an the matche DM Dz k, i.e. the measure isparity, both with their respective variances, we upate the preicte isparity using Kalman filter on the pixel level. As escribe in Section 2.2 we preict the new x an y coorinates of each isparity eterministically, while isparity is estimate by Kalman filtering. First we etermine the Kalman gain: p k k 1 K k = (10) p k k 1 + r k where r k enotes the variance of the matche isparity (measurement) as escribe in Subsection The final isparity is estimate using the stanar Kalman filter upate equation: k = k k 1 + K k ( k z k k 1 ) (11) an the variance is upate as follows p k = (1 K k )p k k 1. (12) Disparity consistency check. Disparity consistency check is performe for two reasons. First, outliers present in the current DM can be propagate to the next frame. Secon, even though propagate isparity is only use to reuce search space, the real minimum of the cost function coul be outsie of the reuce search space. We aress this problem by using several filtering techniques for rejecting outliers like left-right consistency check an sum of absolute ifference (SAD) check on left an right images. Using this metho we reject occlue pixels that are often poorly matche an which introuce outliers being propagate to future frames. In the next frame, for the rejecte pixels, the search is then performe in the whole search space. Please confer Hu an Morohai (2010) for an overview of alternative methos. 3. EXPERIMENTAL VALIDATION To test an valiate the propose algorithm we use the KITTI ataset (Menze an Geiger (2015)), which inclues 200 training an testing scenes. Every scene consist of 20 stereo sequences of real worl images, thus making them appropriate for valiation of algorithms that rely on image sequences rather than a single image frame. Since KITTI ataset is capture on the streets of Karlsruhe, moving objects are present in the majority scenes. In orer to evaluate the propose algorithm, we selecte scenes where no moving objects were present. Current implementation treats the scene as static an any moving objects will introuce preiction errors since their movement is neglecte, thereby resulting with ba final DM estimation. As iscusse in Section 1, this limitation is in fact a base for future ense scene flow estimation an moving object etection. Furthermore, it can also be use in applications where moving objects are not ominant or even non-existent, like space exploration robotics. As escribe in Section 2.3 we use SGM with isparity estimation in the reuce search space base on egomotion. Demonstration of an arbitrary optimize path with matching costs is shown in Fig. 2. Figure 2b shows optimize matching costs of the path shown in Fig. 2a as a prouct of our basic SGM implementation. Next, Fig. 2c highlights the chosen minima, i.e., chosen isparity values on the whole path. The iea of the propose algorithm is to reuce the isparity search space base on vali isparity preictions as shown in Fig. 2. Soli gray color represents the ignore isparity search space. After the search space reuction, minima can be locate faster while reucing the chance of false isparity etection. Figure 2e highlights the reuce search space chosen minima, from which reuction of require computation effort can be seen. In the first iteration, whole isparity search space is consiere since there is no isparity preiction available. In the following frames the isparity preiction is available an SGM can be performe on the reuce isparity search space. Our experiments show that less than 50% of whole isparity search space is consiere in most scenes. For the scene shown in Fig. 2, the consiere isparity search space is 100%, 49.05%, 47.24% an 46.05% for frames 0, 1, 2 an 3, respectively. As the preiction becomes enser the whole computing process becomes faster ue to the reuce search space. Figure 3 emonstrates the whole isparity estimation process. First we take the DM of the previous frame (Fig. 3a) an form a preiction base on ego-motion (Fig. 3b). Here, the zooming effect is easily noticeable. We successfully negate that effect by interpolation metho escribe in Section 2.2 as shown in Fig. 3c. Next, leftright consistency check is performe on left an right DM to compare their results an iscar all isparity preictions that are not below the similarity threshol. This metho is use to reuce a number of outliers cause by incorrect ego-motion propagation or poorly matche isparities from previous frames. The most noticeable rejecte area is visible on the left sie of Fig. 3 where all

5 (a) (a) (b) (b) (c) () (e) (c) Fig. 2: Disparity search space reuction on an optimize matching cost function. (a) shows the left image of stereo frame with highlighte path example, (b) shows optimize matching costs on the path, (c) highlights the chosen minima (isparities) () shows optimize matching costs with the reuce isparity search space (e) highlights the chosen minima (isparities) after search space reuction. the isparities that cannot be observe with both left an right cameras are marke as invali. Fig. 3 also represents the final DM preiction use for further processing. Fig. 3e shows the results of our implementation of an eight-path SGM algorithm with reuce isparity search space. It is noticeable that preicte an matche DM (Figs. 3 an 3e) are very similar. The main ifferences are in the areas of occlue pixels ue to ifferent perspectives of previous an current stereo frames. Other noticeable ifference is near object eges where ege rejection filtering escribe in Section 2.2 is applie after the preiction step. These rejections are necessary to reuce the number of outliers that coul propagate to future frames an result with wrong isparity search interval in the matching process. Finally, we upate the DM as escribe in Section 2.4. The results are shown in Fig. 3f. We aitionally emonstrate the benefits of this approach by referring to Fig. 4, where it is visible that the upate DM combines best isparities from both the matche an preicte DM. As elaborate in Section 2.2, rejecte isparities near isparity iscontinuities can be seen in the preicte DM in Fig. 4. While useful for preserving the eges an shape of foregroun objects, this rejection can result with holes in the preicte backgroun isparities. On the other han, newly matche DM in Fig. 4 shows enser backgroun reconstruction, but is unable to reconstruct foregroun objects properly (the traffic light pole). By using the preicte an matche DM, we can combine best attributes of both reconstructions. This way, a enser map with less outliers is generate. In orer to evaluate the propose approach, we use KITTI benchmarking scripts, where a ba isparity refers to isparities with error grater than 3 pixels or 5% relative to the groun truth isparities acquire by the 3D laser range sensor. Direct output of the evaluation for four ifferent scenes is shown in Fig. 5. As Table 1 shows, our approach results with less ba isparities than the classical eight-path SGM, while consiering less than 50% of the whole isparity search space. This way we increase both computation spee an accuracy of DM estimation. () (e) (f) Fig. 3: Demonstration of isparity processing. Images show left DM of (a) previous frame, (b) preiction of current frame base on ego-motion an previous DM, (c) preiction of current frame interpolate ue to zooming effect, () preiction of current frame filtere with leftright consistency check, (e) SGM matching result on reuce search isparity space, (f) upate () with (e). Fig. 4: The comparison of (1) matche DM, (2) preicte DM, (3) upate DM an (4) intensity image. Our experiments inicate that this also implies about 50% reuction of execution time when compare to base SGM.

6 Fig. 5: Direct output of KITTI benchmarking script for scene flow training scenes 130, 84, 87 an 146 respectively (re - ba isparity, blue - goo isparity). Scene SGM Propose Propose, interpolate % 10.19% 4.33% % 11.79% 6.34 % % 12.24% 6.51% % 9.75 % 5.17 % Table 1: Comparison of the propose an basic SGM stereo matching accuracy on the KITTI benchmark scenes CONCLUSION AND FURTHER WORK In this paper we have presente a stable, accurate, an efficient spatio-temporal isparity estimation algorithm. The propose approach is base on using ego-motion between consecutive frames to transform the DM from the previous frame to the current one. The transforme isparity, i.e., the preicte isparity, is use as a reference in the isparity search space, which is reuce base on the preicte isparity uncertainty. The newly matche isparity from the reuce search space is then use as a measurement within the Kalman filter in orer to upate the preicte isparity. This results with reuce computational effort an an increase in the accuracy of estimate DM compare to the basic SGM. We constraine our use-case to scenes without moving objects, since our primary aim was to achieve accurate reconstruction of the static parts of the scene, which will then serve as a base for ense scene flow estimation an moving object etection. Moreover, the propose approach can be use in applications with no moving objects, e.g., in space exploration robotics. We teste the algorithm on the KITTI benchmark an shown that it can achieve better accuracy than the basic SGM implementation, while reucing the isparity search space. REFERENCES Cvišić, I. an Petrović, I. (2015). Stereo oometry base on careful feature selection an tracking. European Conference on Mobile Robots (ECMR). Demirjian, D. an Darrell, T. (2001). Motion estimation from isparity images. In IEEE International Conference on Computer Vision (ICCV). Dobias, M. an Sara, R. (2011). Real-time global preiction for temporally stable stereo. In IEEE International Conference on Computer Vision (ICCV), Franke, U., Rabe, C., Baino, H., an Gehrig, S. (2005). 6D-Vision : Fusion of Stereo an Motion for Robust Environment Perception. Lecture Notes in Computer Science, 3663, Hirschmüller, H. (2008). Stereo Processing by Semi-global Matching an Mutual Information. IEEE Transactions on Pattern Analysis an Machine Intelligence, 30(2), Hu, X. an Morohai, P. (2010). Evaluation of stereo confience inoors an outoors. In Proceeings of the IEEE Computer Society Conference on Computer Vision an Pattern Recognition, Jakob Engel, Jörg Stückler, D.C. (2015). Large-Scale Direct SLAM with Stereo Cameras. In International Conference on Intelligent Robots an Systems (IROS). Jiang, J., Cheng, J., Chen, B., an Wu, X. (2014). Disparity preiction between ajacent frames for ynamic scenes. Neurocomputing, 142, Klaus, A., Sormann, M., an Karner, K. (2006). Segment- Base Stereo Matching Using Belief Propagation an a Self-Aapting. In International Conference on Pattern Recognition (ICPR). Kolmogorov, V. an Zabih, R. (2001). Computing visual corresponences with occlusions using graph cuts. In International Conference for Computer Vision, Matthies, L., Kanae, T., an Szeliski, R. (1989). Kalman Filter-base algorithms for estimating epth from image sequences. International Journal of Computer Vision, 3, Mayer, N., Ilg, E., Häusser, P., Fischer, P., Cremers, D., Dosovitskiy, A., an Brox, T. (2016). A Large Dataset to Train Convolutional Networks for Disparity, Optical Flow, an Scene Flow Estimation. In IEEE International Conference on Computer Vision an Pattern Recognition (CVPR). Menze, M. an Geiger, A. (2015). Object scene flow for autonomous vehicles. In Conference on Computer Vision an Pattern Recognition (CVPR). Morales, S. an Klette, R. (2013). Kalman-filter base spatio-temporal isparity integration. Pattern Recognition Letters, 34(8), Vaurey, T., Baino, H., an Gehrig, S. (2008). Integrating isparity images by incorporating isparity rate. Lecture Notes in Computer Science, 4931, Weel, A., Brox, T., Vaurey, T., Rabe, C., Franke, U., an Cremers, D. (2011). Stereoscopic Scene Flow Computation for 3D Motion Unerstaning. International Journal of Computer Vision, 95(1), Yang, Q., Wang, L., Yang, R., Stewénius, H., an Nistér, D. (2009). Stereo matching with color-weighte correlation, hierarchical belief propagation, an occlusion hanling. IEEE Transactions on Pattern Analysis an Machine Intelligence, 31(3), Žbontar, J. an LeCun, Y. (2015). Stereo Matching by Training a Convolutional Neural Network to Compare Image Patches. In Conference on Computer Vision an Pattern Recognition (CVPR), 2002.