Depth Mapping Algorithm Performance Analysis
Introduction
We will implement various disparity estimation algorithms and compare their performance. ADD MORE
Background
Disparity and Depth
Depth information about a scene can be captured using a stereo camera (2 cameras that are separated horizontally but aligned vertically). The stereo image pair taken by the stereo camera contains this depth information in the horizontal differences (when comparing the stereo image pair, objects closer to the camera will be more horizontally displaced). These differences (also called disparities) can be used to determine the relative distance from the camera for different objects in the scene. In Figure 1, you can see such differences on the left where the red and blue don't match up.
Disparity and depth can be related by the following equation (where x-x' is disparity, z is depth, f is the focal length, and B is the interocular distance).
!!!!!!!!!!!!OSCAR/SHALINI WRITE STUFF HERE!!!!!!!!!!!!!!!!!!!!!! Since depth reduced to a scaled quantity of disparity, the accuracy of the disparity could be said to be the more important calculation and disparity calculation could be interchangeable with depth calculation.
Image Rectification
In order to extract depth information, the stereo image pair must first be rectified (i.e. the images must be transformed in some way such that the only differences that remain are horizontal differences corresponding to the distance of the object from the camera). Rectification can be accomplished both with and without camera calibration. If the corresponding camera intrinsics and extrinsics are given for the stereo image pair, then calibration is not necessary. If they are not given, but photos of a checkerboard or some other calibration object are provided, then the calibration parameters can be calculated.
If not enough camera parameters are given and there are no checkerboard images to be used for calibration, then the following 4 steps can be used for rectification of a given stereo image pair.
- First we detect SURF keypoints (a scale and rotation invariant feature detector) in each stereo image and extract the feature vectors of each keypoint.
- We then find matching keypoints between the images using a similarity metric. The uncalibrated rectification used by MATLAB uses the sum of absolute differences metric.
- We then remove outliers (incorrect matches) using an epipolar constraint. Looking at Figure 3, this means that for a keypoint x on the left image, the matching keypoint on the right image must lie on the corresponding epipolar line defined by the intersection of the epipolar plane and the image plane.
- Now using the vector distance between the remaining correct matches, we must compute a 2D projective geometric transformation to apply to one or both of the stereo images. After transformation, the stereo images should no longer have any vertical displacement.
Datasets
SYNS Dataset
The Southampton-York Natural Scenes (SYNS) dataset contains image and 3D range data measured from different rural and urban locations. Each sample contains the following data:
- LiDAR depth information (360 x 135 degree field of view)
- Panoramic HDR image captured with a SpheroCam (Nikkor fish-eye lens) (360 x 180 degree image)
- Stereo image pairs captured with 2 Nikon DSLR cameras (each image pair was captured at a different rotation of the camera such that all of them covered a 360 degree view of the surroundings)
This dataset did not include any intrinsics or extrinsics, or even any calibration photos from which we could extract this data. Therefore, we used the uncalibrated rectification method described above on the stereo images in order to compute disparity maps. An example of a rectified stereo image pair is shown below in Figure 4 (left). We then used MATLAB's disparity mapping algorithm on these rectified stereo images to generate the disparity map shown below in Figure 4 (right).
Our goals for this dataset were to compute depth map using various algorithms (described in the Methods section below) and then compare the results to LiDAR info (ground truth). In order to do a quantitative evaluation of this comparison, we would need to match the stereo view to the exact matching region of interest in the LiDAR data. We would be able to create a projection of the pointcloud LiDAR data if we have information about the position and angle of the camera relative to the scene. However, this information was not provided, so we would have to instead use a more brute force method to determine which area of the LiDAR data corresponds to the stereo view being considered. To do this we would have to iterate over the HDR image and quantitatively compare it to the stereo images using a method such as least mean squares. We utilized a MATLAB script to scale the panoramic LiDAR info to the panoramic HDR image, so that we can find the LiDAR information that corresponds to the region of interest found in the HDR image. A sample output of this script is shown in Figure 5. We can then extract this relevant LiDAR info and compare it to the computed depth map.
Talk about obstacles.
Middlebury Dataset
The Middlebury dataset was created by Nera Nesic, Porter Westling, Xi Wang, York Kitajima, Greg Krathwohl, and Daniel Scharstein at Middlebury College during 2011-2013, and refined with Heiko Hirschmüller at the DLR Germany during 2014[2]. It is comprised of 33 stereo image datasets, 20 of them are taken using the new Middlebury Stereo Evaluation (10 each for training and test sets). The detailed description of the acquisition process can be found in our GCPR 2014 paper [4].
This dataset is highly cited and provides details about all the intrinsic and extrinsic parameters of the camera used, as well as images with "perfect" and "imperfect" rectification. It does not have a "ground truth" depth as with the SYNS LIDAR data, however it provides a highly accurate "ground truth" disparity map for its datasets which were acquired using through multiple iterations and refinement details in the Figure 7 below:
All this information and set-up helped bypass the issues with the SYNS dataset. The "perfectly" rectified stereo images from the dataset were used easy the process and see direct comparisons between methods mentioned in the later sections.
Methods: Block Matching and Similarity Metrics
To extract the disparity between two stereo images there are 4 methods which were compared. All of these methods utilize block/window matching to produce their results.
Block/Window Matching
To analyse an image it is broken down into subsections called blocks or windows, and it is these blocks which comparisons are performed on.
 |
 |
SAD
The sum of absolute differences (SAD) is a measure of the similarity between image blocks. It is calculated by taking the absolute difference between each pixel in the original block and the corresponding pixel in the block being used for comparison. [3]
- Compute the difference and sum over all pixels
SSD
The sum of squared difference takes the squared differences within the two blocks. This measure has a higher computational complexity compared to SAD algorithm as it involves numerous multiplication operations. [3]
- Compute the squared difference and sum over all pixels
NCC
Normalized cross-correlation is a metric to evaluate the degree of similarity (or dissimilarity) between two compared images. In NCC, a window of suitable size is determined and moved over the entire image or the cost matrix. The correspondence is thus obtained by dividing the normalized summation of the product of intensities over the entire window by the standard deviation of the intensities of the images over the entire window. [3]
- Compute the cross-correlation
CT
The census transform encodes the relative brightness of each pixel (with respect to its neighbors) and compares the digital encoding of pixels from windows from each image (left and right). The algorithm is as follows:
- Compute bit-string for pixel p, based on intensities of neighboring pixels
- Compare left image bit-string with bit-strings from a range of windows in the right image
- Choose disparity d with lowest Hamming distance
Methods: Algorithm Evaluation
Talk about
- Tried these 3 algorithms, x pictures for each
- How error rate is calculated and/or averaged
- Other stuff
Results
Sum of Squared Differences
Sum of Absolute Difference
Performance with default parameters
Effect of Block Size and Smoothing
For block matching with and without the semi-global smoothing, we tested the effect of altering the block size of the Block Matching + SAD algorithm. The chart below shows the resulting average error rates over 15 images, along with the corresponding graph.
Overall, the semi-global smoothing had a significantly lower error rate than regular SAD. Furthermore, the block size affects the performance of each differently -- semi-global SAD produces lower error rates with smaller block sizes, and SAD has a local optimum around a block size of 11.
Comparing the computational time of each algorithm over all block sizes tested, SAD with semi-global smoothing took an average of 0.9598 seconds per image, while SAD took an average of 0.3016 seconds. Thus, regular SAD is more than three times as fast.
Census Transformation
Conclusions
Available Datasets
Running algorithms and experiments with the SYNS and Middlebury datasets highlighted the tradeoffs between the different stereo image and depth-related datasets that currently exist. SYNS provides a fantastic range of practical real-world scenes, and even includes the ideal ground truth LiDAR information for each scene. However, the lack of camera intrinsics and extrinsics or even just checkerboard images from which we could extract the calibration parameters proved to be prohibitive when trying to map the stereo images to the LiDAR information. This makes the dataset very difficult to use for depth mapping projects.
The Middlebury dataset is currently one of the most widely used datasets for depth mapping, since it has thoroughly documented calibration information as well as images rectified using multiple different methods. It also has a highly accurate disparity map to be used as a ground truth when running our own depth mapping algorithms. However, it would be more ideal if this dataset had true depth information (like LiDAR) that could be used as the ideal ground truth.
Our conclusion with regards to the datasets is that a new ideal dataset could be created that is comprehensive in both its documentation of calibration details and scene capture setup, but also includes true depth and disparity information for comparative studies.
Smoothing in Disparity Algorithms
From the comparison of the SAD algorithm with and without semi-global smoothing, it appears that overall, smoothing increases the accuracy of the disparity map. This is most likely due to the fact that in block matching, it is difficult to identify the disparity between regions of uniform intensity. Thus, uniform regions contain more noise and inaccuracies in the final disparity map. By forcing similar disparity on neighboring blocks, these regions reflect the actual disparity more closely. Furthermore, as shown in Figure 6, while small block size increased the error of regular SAD, the error rate of SAD with smoothing consistently decreased with smaller block size. This may indicate that block matching may not work very well with the limited scope of a small block, but the smoothing reduces the errors caused by this.
Which Algorithm Performs 'Best'?
In general, the performance of the Census Tract algorithm was best when tested on the 15 training images in the Middlebury dataset, with a 20% error rate. However, it took an average of 42.9 seconds to run on each image. On the other hand, SAD took, on average, only 0.3 seconds per image. Thus, if one has a time constraint, using block matching with SAD or SSD may be better suited. SSD is more sensitive to differences in blocks, so may be more accurate when computing disparity, but the squaring operation is more computationally expensive than the absolute value operation in SAD. Thus, the optimal algorithm depends on the demands of one's application.
References
[1] Daniel Scharstein, et. al., “High-Resolution Stereo Datasets with Subpixel-Accurate Ground Truth” German Conference on Pattern Recognition. 2014. http://www.cs.middlebury.edu/~schar/papers/datasets-gcpr2014.pdf
[2] Jongchul Lee, et. al., "Improved census transform for noise robust stereo matching," Optical Engineering 55(6), 063107 (16 June 2016). http://dx.doi.org/10.1117/1.OE.55.6.063107
[3] Pranoti Dhole, et. al., "Depth Map Estimation Using SIMULINK Tool," 3rd International Conference on Signal Processing and Integrated Networks. 2016.http://ieeexplore.ieee.org/stamp/stamp.jsp?arnumber=7566714
[4] D. Scharstein, H. Hirschmüller, Y. Kitajima, G. Krathwohl, N. Nesic, X. Wang, and P. Westling. High-resolution stereo datasets with subpixel-accurate ground truth. In German Conference on Pattern Recognition (GCPR 2014), Münster, Germany, September 2014.
[5]Chris McCormick, Stereo Vision Tutorial - Part I. 2014. http://mccormickml.com/2014/01/10/stereo-vision-tutorial-part-i/
Appendix I
- MATLAB code for disparity map algorithms and evaluation code: File:Projectcode.zip
- Link to Middlebury dataset: http://vision.middlebury.edu/stereo/data/scenes2014/
- Link to SYNS dataset: https://syns.soton.ac.uk/
Appendix II
We divided the work as follows:
- Oscar Guerrero:
- Deepti Mahajan: Researched and implemented disparity algorithm code, ran experiments and gathered results, worked on presentation and wiki
- Shalini Ranmuthu: Implemented algorithms for SYNS dataset, researched and implemented image rectification methods, worked on presentation and wiki