OpenCV has many algorithms covering the same subject. When implementing a new processing pipeline, sometimes there is more than one choice for a step in the pipeline. For example, in Chapter 14, Explore Structure from Motion with the SfM Module, we made an arbitrary decision to use AKAZE features for finding landmarks between the images to estimate camera motion, and sparse 3D structure, however; there are many more kinds of 2D features available in OpenCV's features2D module. A more sensible mode of operation should have been to select the type of feature algorithm to use based on its performance, with respect to our needs. At the very least, we need to be aware of the different options.
Again, we looked to create a convenient way to see whether there are multiple options for the same task. We created a table where we list specific computer vision tasks that have multiple algorithm implementations in OpenCV. We also strived to mark whether algorithms have a common abstract API, and thus, easily and completely interchangeable within the code. While OpenCV offers the cv::Algorithm base class abstraction for most if not all of its algorithms, the abstraction is at a very high level and gives very little power to polymorphism and interchangeability. From our review, we exclude the machine learning algorithms (the ml module and the cv::StatsModel common API) since they are not proper computer vision algorithms, as well as low-level image processing algorithms, which do in fact have overlapping implementations (for example, the Hough detector family). We also exclude the GPU CUDA implementations that shadow several core topics such as object detection, background segmentation, 2D features, and more, since they are mostly replicas of the CPU implementations.
The following are topics with multiple implementations in OpenCV:
| Topic | Implementations | Base API? |
|
Optical flow |
video module: SparsePyrLKOpticalFlow, FarnebackOpticalFlow, DISOpticalFlow, VariationalRefinement optflow contrib module: DualTVL1OpticalFlow, OpticalFlowPCAFlow |
Yes |
|
Object tracking |
track contrib module: TrackerBoosting, TrackerCSRT, TrackerGOTURN, TrackerKCF, TrackerMedianFlow, TrackerMIL, TrackerMOSSE, TrackerTLD External: DetectionBasedTracker |
Yes1 |
|
Object detection |
objdetect module: CascadeClassifier, HOGDescriptor, QRCodeDetector, linemod contrib module: Detector aruco contrib module: aruco::detectMarkers |
No2
|
|
2D features |
OpenCV's most established common API. features2D module: AgastFeatureDetector, AKAZE, BRISK, FastFeatureDetector, GFTTDetector, KAZE, MSER, ORB, SimpleBlobDetector xfeatures2D contrib module: BoostDesc, BriefDescriptorExtractor, DAISY, FREAK, HarrisLaplaceFeatureDetector, LATCH, LUCID, MSDDetector, SIFT, StarDetector, SURF, VGG |
Yes |
|
Feature matching |
BFMatcher, FlannBasedMatcher |
Yes |
|
Background subtraction |
video module: BackgroundSubtractorKNN, BackgroundSubtractorMOG2 bgsegm contrib module: BackgroundSubtractorCNT, BackgroundSubtractorGMG, BackgroundSubtractorGSOC, BackgroundSubtractorLSBP, BackgroundSubtractorMOG |
Yes |
|
Camera calibration |
calib3d module: calibrateCamera, calibrateCameraRO, stereoCalibrate aruco contrib module: calibrateCameraArcuo, calibrateCameraCharuco ccalib contrib module: omnidir::calibrate, omnidir::stereoCalibrate |
No |
|
Stereo reconstruction |
calib3d module: StereoBM, StereoSGBM stereo contrib module: StereoBinaryBM, StereoBinarySGBM ccalib contrib module: omnidir::stereoReconstruct |
Partial3 |
|
Pose estimation |
solveP3P, solvePnP, solvePnPRansac |
No |
1 Only for the classes in the track contrib module.
2 Some classes share functions with the same name, but no inherited abstract class.
3 Each module has a base within itself, but not shared across modules.
When approaching a problem with a few algorithmic options, it's important not to commit too early to one execution path. We may use the preceding table above to see options exist, and then explore them. Next, we will discuss how to select from a pool of options.