# Writing a new PCL class

Converting code to a PCL-like mentality/syntax for someone that comes in contact for the first time with our infrastructure might appear difficult, or raise certain questions.

This short guide is to serve as both a HowTo and a FAQ for writing new PCL classes, either from scratch, or by adapting old code.

Besides converting your code, this guide also explains some of the advantages of contributing your code to an already existing open source project. Here, we advocate for PCL, but you can certainly apply the same ideology to other similar projects.

The first question that someone might ask and we would like to answer is:

Why contribute to PCL, as in what are its advantages?

This question assumes you’ve already identified that the set of tools and libraries that PCL has to offer are useful for your project, so you have already become an user.

Because open source projects are mostly voluntary efforts, usually with developers geographically distributed around the world, it’s very common that the development process has a certain incremental, and iterative flavor. This means that:

• it’s impossible for developers to think ahead of all the possible uses a new piece of code they write might have, but also…
• figuring out solutions for corner cases and applications where bugs might occur is hard, and might not be desirable to tackle at the beginning, due to limited resources (mostly a cost function of free time).

In both cases, everyone has definitely encountered situations where either an algorithm/method that they need is missing, or an existing one is buggy. Therefore the next natural step is obvious:

change the existing code to fit your application/problem.

While we’re going to discuss how to do that in the next sections, we would still like to provide an answer for the first question that we raised, namely “why contribute?”.

In our opinion, there are many advantages. To quote Eric Raymond’s Linus’s Law: “given enough eyeballs, all bugs are shallow”. What this means is that by opening your code to the world, and allowing others to see it, the chances of it getting fixed and optimized are higher, especially in the presence of a dynamic community such as the one that PCL has.

In addition to the above, your contribution might enable, amongst many things:

• others to create new work based on your code;
• you to learn about new uses (e.g., thinks that you haven’t thought it could be used when you designed it);
• worry-free maintainership (e.g., you can go away for some time, and then return and see your code still working. Others will take care of adapting it to the newest platforms, newest compilers, etc);
• your reputation in the community to grow - everyone likes free stuff (!).

For most of us, all of the above apply. For others, only some (your mileage might vary).

# Example: a bilateral filter

To illustrate the code conversion process, we selected the following example: apply a bilateral filter over intensity data from a given input point cloud, and save the results to disk.

 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 #include #include #include typedef pcl::PointXYZI PointT; float G (float x, float sigma) { return exp (- (x*x)/(2*sigma*sigma)); } int main (int argc, char *argv[]) { std::string incloudfile = argv[1]; std::string outcloudfile = argv[2]; float sigma_s = atof (argv[3]); float sigma_r = atof (argv[4]); // Load cloud pcl::PointCloud::Ptr cloud (new pcl::PointCloud); pcl::io::loadPCDFile (incloudfile.c_str (), *cloud); int pnumber = (int)cloud->size (); // Output Cloud = Input Cloud pcl::PointCloud outcloud = *cloud; // Set up KDTree pcl::KdTreeFLANN::Ptr tree (new pcl::KdTreeFLANN); tree->setInputCloud (cloud); // Neighbors containers std::vector k_indices; std::vector k_distances; // Main Loop for (int point_id = 0; point_id < pnumber; ++point_id) { float BF = 0; float W = 0; tree->radiusSearch (point_id, 2 * sigma_s, k_indices, k_distances); // For each neighbor for (size_t n_id = 0; n_id < k_indices.size (); ++n_id) { float id = k_indices.at (n_id); float dist = sqrt (k_distances.at (n_id)); float intensity_dist = abs (cloud->points[point_id].intensity - cloud->points[id].intensity); float w_a = G (dist, sigma_s); float w_b = G (intensity_dist, sigma_r); float weight = w_a * w_b; BF += weight * cloud->points[id].intensity; W += weight; } outcloud.points[point_id].intensity = BF / W; } // Save filtered output pcl::io::savePCDFile (outcloudfile.c_str (), outcloud); return (0); }
The presented code snippet contains:
• an I/O component: lines 21-27 (reading data from disk), and 64 (writing data to disk)
• an initialization component: lines 29-35 (setting up a search method for nearest neighbors using a KdTree)
• the actual algorithmic component: lines 7-11 and 37-61

Our goal here is to convert the algorithm given into an useful PCL class so that it can be reused elsewhere.

# Setting up the structure

Note

If you’re not familiar with the PCL file structure already, please go ahead and read the PCL C++ Programming Style Guide to familiarize yourself with the concepts.

There’s two different ways we could set up the structure: i) set up the code separately, as a standalone PCL class, but outside of the PCL code tree; or ii) set up the files directly in the PCL code tree. Since our assumption is that the end result will be contributed back to PCL, it’s best to concentrate on the latter, also because it is a bit more complex (i.e., it involves a few additional steps). You can obviously repeat these steps with the former case as well, with the exception that you don’t need the files copied in the PCL tree, nor you need the fancier cmake logic.

Assuming that we want the new algorithm to be part of the PCL Filtering library, we will begin by creating 3 different files under filters:

• include/pcl/filters/bilateral.h - will contain all definitions;
• include/pcl/filters/impl/bilateral.hpp - will contain the templated implementations;
• src/bilateral.cpp - will contain the explicit template instantiations [*].

We also need a name for our new class. Let’s call it BilateralFilter.

 [*] Some PCL filter algorithms provide two implementations: one for PointCloud types and another one operating on legacy PCLPointCloud2 types. This is no longer required.

## bilateral.h

As previously mentioned, the bilateral.h header file will contain all the definitions pertinent to the BilateralFilter class. Here’s a minimal skeleton:

 1 2 3 4 5 6 7 8 9 10 11 12 13 14 #ifndef PCL_FILTERS_BILATERAL_H_ #define PCL_FILTERS_BILATERAL_H_ #include namespace pcl { template class BilateralFilter : public Filter { }; } #endif // PCL_FILTERS_BILATERAL_H_

## bilateral.hpp

While we’re at it, let’s set up two skeleton bilateral.hpp and bilateral.cpp files as well. First, bilateral.hpp:

 1 2 3 4 5 6 #ifndef PCL_FILTERS_BILATERAL_IMPL_H_ #define PCL_FILTERS_BILATERAL_IMPL_H_ #include #endif // PCL_FILTERS_BILATERAL_H_

This should be straightforward. We haven’t declared any methods for BilateralFilter yet, therefore there is no implementation.

## bilateral.cpp

Let’s write bilateral.cpp too:

 1 2 #include #include

Because we are writing templated code in PCL (1.x) where the template parameter is a point type (see Adding your own custom PointT type), we want to explicitly instantiate the most common use cases in bilateral.cpp, so that users don’t have to spend extra cycles when compiling code that uses our BilateralFilter. To do this, we need to access both the header (bilateral.h) and the implementations (bilateral.hpp).

## CMakeLists.txt

Let’s add all the files to the PCL Filtering CMakeLists.txt file, so we can enable the build.

 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 # Find "set (srcs", and add a new entry there, e.g., set (srcs src/conditional_removal.cpp # ... src/bilateral.cpp) ) # Find "set (incs", and add a new entry there, e.g., set (incs include pcl/${SUBSYS_NAME}/conditional_removal.h # ... include pcl/${SUBSYS_NAME}/bilateral.h ) # Find "set (impl_incs", and add a new entry there, e.g., set (impl_incs include/pcl/${SUBSYS_NAME}/impl/conditional_removal.hpp # ... include/pcl/${SUBSYS_NAME}/impl/bilateral.hpp )

# Filling in the class structure

If you correctly edited all the files above, recompiling PCL using the new filter classes in place should work without problems. In this section, we’ll begin filling in the actual code in each file. Let’s start with the bilateral.cpp file, as its content is the shortest.

## bilateral.cpp

As previously mentioned, we’re going to explicitly instantiate and precompile a number of templated specializations for the BilateralFilter class. While this might lead to an increased compilation time for the PCL Filtering library, it will save users the pain of processing and compiling the templates on their end, when they use the class in code they write. The simplest possible way to do this would be to declare each instance that we want to precompile by hand in the bilateral.cpp file as follows:

 1 2 3 4 5 6 7 8 #include #include #include template class PCL_EXPORTS pcl::BilateralFilter; template class PCL_EXPORTS pcl::BilateralFilter; template class PCL_EXPORTS pcl::BilateralFilter; // ...

However, this becomes cumbersome really fast, as the number of point types PCL supports grows. Maintaining this list up to date in multiple files in PCL is also painful. Therefore, we are going to use a special macro called PCL_INSTANTIATE and change the above code as follows:

 1 2 3 4 5 6 #include #include #include #include PCL_INSTANTIATE(BilateralFilter, PCL_XYZ_POINT_TYPES);

This example, will instantiate a BilateralFilter for all XYZ point types defined in the point_types.h file (see PCL_XYZ_POINT_TYPES for more information).

By looking closer at the code presented in Example: a bilateral filter, we notice constructs such as cloud->points[point_id].intensity. This indicates that our filter expects the presence of an intensity field in the point type. Because of this, using PCL_XYZ_POINT_TYPES won’t work, as not all the types defined there have intensity data present. In fact, it’s easy to notice that only two of the types contain intensity, namely: PointXYZI and PointXYZINormal. We therefore replace PCL_XYZ_POINT_TYPES and the final bilateral.cpp file becomes:

 1 2 3 4 5 6 #include #include #include #include PCL_INSTANTIATE(BilateralFilter, (pcl::PointXYZI)(pcl::PointXYZINormal));

Note that at this point we haven’t declared the PCL_INSTANTIATE template for BilateralFilter, nor did we actually implement the pure virtual functions in the abstract class pcl::Filter so attemping to compile the code will result in errors like:

filters/src/bilateral.cpp:6:32: error: expected constructor, destructor, or type conversion before ‘(’ token

## bilateral.h

We begin filling the BilateralFilter class by first declaring the constructor, and its member variables. Because the bilateral filtering algorithm has two parameters, we will store these as class members, and implement setters and getters for them, to be compatible with the PCL 1.x API paradigms.

 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 ... namespace pcl { template class BilateralFilter : public Filter { public: BilateralFilter () : sigma_s_ (0), sigma_r_ (std::numeric_limits::max ()) { } void setSigmaS (const double sigma_s) { sigma_s_ = sigma_s; } double getSigmaS () { return (sigma_s_); } void setSigmaR (const double sigma_r) { sigma_r_ = sigma_r; } double getSigmaR () { return (sigma_r_); } private: double sigma_s_; double sigma_r_; }; } #endif // PCL_FILTERS_BILATERAL_H_

Nothing out of the ordinary so far, except maybe lines 8-9, where we gave some default values to the two parameters. Because our class inherits from pcl::Filter, and that inherits from pcl::PCLBase, we can make use of the setInputCloud method to pass the input data to our algorithm (stored as input_). We therefore add an using declaration as follows:

 1 2 3 4 5 6 7 8 ... template class BilateralFilter : public Filter { using Filter::input_; public: BilateralFilter () : sigma_s_ (0), ...

This will make sure that our class has access to the member variable input_ without typing the entire construct. Next, we observe that each class that inherits from pcl::Filter must inherit a applyFilter method. We therefore define:

 1 2 3 4 5 6 7 8 9 10 11 12 13 ... using Filter::input_; typedef typename Filter::PointCloud PointCloud; public: BilateralFilter () : sigma_s_ (0), sigma_r_ (std::numeric_limits::max ()) { } void applyFilter (PointCloud &output); ...

The implementation of applyFilter will be given in the bilateral.hpp file later. Line 3 constructs a typedef so that we can use the type PointCloud without typing the entire construct.

Looking at the original code from section Example: a bilateral filter, we notice that the algorithm consists of applying the same operation to every point in the cloud. To keep the applyFilter call clean, we therefore define method called computePointWeight whose implementation will contain the corpus defined in between lines 45-58:

 1 2 3 4 5 6 7 ... void applyFilter (PointCloud &output); double computePointWeight (const int pid, const std::vector &indices, const std::vector &distances); ...

In addition, we notice that lines 29-31 and 43 from section Example: a bilateral filter construct a KdTree structure for obtaining the nearest neighbors for a given point. We therefore add:

 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 #include ... using Filter::input_; typedef typename Filter::PointCloud PointCloud; typedef typename pcl::KdTree::Ptr KdTreePtr; public: ... void setSearchMethod (const KdTreePtr &tree) { tree_ = tree; } private: ... KdTreePtr tree_; ...

Finally, we would like to add the kernel method (G (float x, float sigma)) inline so that we speed up the computation of the filter. Because the method is only useful within the context of the algorithm, we will make it private. The header file becomes:

 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 #ifndef PCL_FILTERS_BILATERAL_H_ #define PCL_FILTERS_BILATERAL_H_ #include #include namespace pcl { template class BilateralFilter : public Filter { using Filter::input_; typedef typename Filter::PointCloud PointCloud; typedef typename pcl::KdTree::Ptr KdTreePtr; public: BilateralFilter () : sigma_s_ (0), sigma_r_ (std::numeric_limits::max ()) { } void applyFilter (PointCloud &output); double computePointWeight (const int pid, const std::vector &indices, const std::vector &distances); void setSigmaS (const double sigma_s) { sigma_s_ = sigma_s; } double getSigmaS () { return (sigma_s_); } void setSigmaR (const double sigma_r) { sigma_r_ = sigma_r; } double getSigmaR () { return (sigma_r_); } void setSearchMethod (const KdTreePtr &tree) { tree_ = tree; } private: inline double kernel (double x, double sigma) { return (exp (- (x*x)/(2*sigma*sigma))); } double sigma_s_; double sigma_r_; KdTreePtr tree_; }; } #endif // PCL_FILTERS_BILATERAL_H_

## bilateral.hpp

There’s two methods that we need to implement here, namely applyFilter and computePointWeight.

 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 template double pcl::BilateralFilter::computePointWeight (const int pid, const std::vector &indices, const std::vector &distances) { double BF = 0, W = 0; // For each neighbor for (size_t n_id = 0; n_id < indices.size (); ++n_id) { double id = indices[n_id]; double dist = std::sqrt (distances[n_id]); double intensity_dist = abs (input_->points[pid].intensity - input_->points[id].intensity); double weight = kernel (dist, sigma_s_) * kernel (intensity_dist, sigma_r_); BF += weight * input_->points[id].intensity; W += weight; } return (BF / W); } template void pcl::BilateralFilter::applyFilter (PointCloud &output) { tree_->setInputCloud (input_); std::vector k_indices; std::vector k_distances; output = *input_; for (size_t point_id = 0; point_id < input_->points.size (); ++point_id) { tree_->radiusSearch (point_id, sigma_s_ * 2, k_indices, k_distances); output.points[point_id].intensity = computePointWeight (point_id, k_indices, k_distances); } }

The computePointWeight method should be straightforward as it’s almost identical to lines 45-58 from section Example: a bilateral filter. We basically pass in a point index that we want to compute the intensity weight for, and a set of neighboring points with distances.

In applyFilter, we first set the input data in the tree, copy all the input data into the output, and then proceed at computing the new weighted point intensities.

Looking back at Filling in the class structure, it’s now time to declare the PCL_INSTANTIATE entry for the class:

 1 2 3 4 5 6 7 8 9 10 #ifndef PCL_FILTERS_BILATERAL_IMPL_H_ #define PCL_FILTERS_BILATERAL_IMPL_H_ #include ... #define PCL_INSTANTIATE_BilateralFilter(T) template class PCL_EXPORTS pcl::BilateralFilter; #endif // PCL_FILTERS_BILATERAL_H_

One additional thing that we can do is error checking on:

• whether the two sigma_s_ and sigma_r_ parameters have been given;
• whether the search method object (i.e., tree_) has been set.

For the former, we’re going to check the value of sigma_s_, which was set to a default of 0, and has a critical importance for the behavior of the algorithm (it basically defines the size of the support region). Therefore, if at the execution of the code, its value is still 0, we will print an error using the PCL_ERROR macro, and return.

In the case of the search method, we can either do the same, or be clever and provide a default option for the user. The best default options are:

• use an organized search method via pcl::OrganizedNeighbor if the point cloud is organized;
• use a general purpose kdtree via pcl::KdTreeFLANN if the point cloud is unorganized.
 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 #include #include ... template void pcl::BilateralFilter::applyFilter (PointCloud &output) { if (sigma_s_ == 0) { PCL_ERROR ("[pcl::BilateralFilter::applyFilter] Need a sigma_s value given before continuing.\n"); return; } if (!tree_) { if (input_->isOrganized ()) tree_.reset (new pcl::OrganizedNeighbor ()); else tree_.reset (new pcl::KdTreeFLANN (false)); } tree_->setInputCloud (input_); ...

The implementation file header thus becomes:

 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 #ifndef PCL_FILTERS_BILATERAL_IMPL_H_ #define PCL_FILTERS_BILATERAL_IMPL_H_ #include #include #include template double pcl::BilateralFilter::computePointWeight (const int pid, const std::vector &indices, const std::vector &distances) { double BF = 0, W = 0; // For each neighbor for (size_t n_id = 0; n_id < indices.size (); ++n_id) { double id = indices[n_id]; double dist = std::sqrt (distances[n_id]); double intensity_dist = abs (input_->points[pid].intensity - input_->points[id].intensity); double weight = kernel (dist, sigma_s_) * kernel (intensity_dist, sigma_r_); BF += weight * input_->points[id].intensity; W += weight; } return (BF / W); } template void pcl::BilateralFilter::applyFilter (PointCloud &output) { if (sigma_s_ == 0) { PCL_ERROR ("[pcl::BilateralFilter::applyFilter] Need a sigma_s value given before continuing.\n"); return; } if (!tree_) { if (input_->isOrganized ()) tree_.reset (new pcl::OrganizedNeighbor ()); else tree_.reset (new pcl::KdTreeFLANN (false)); } tree_->setInputCloud (input_); std::vector k_indices; std::vector k_distances; output = *input_; for (size_t point_id = 0; point_id < input_->points.size (); ++point_id) { tree_->radiusSearch (point_id, sigma_s_ * 2, k_indices, k_distances); output.points[point_id].intensity = computePointWeight (point_id, k_indices, k_distances); } } #define PCL_INSTANTIATE_BilateralFilter(T) template class PCL_EXPORTS pcl::BilateralFilter; #endif // PCL_FILTERS_BILATERAL_H_

# Taking advantage of other PCL concepts

## Point indices

The standard way of passing point cloud data into PCL algorithms is via setInputCloud calls. In addition, PCL also defines a way to define a region of interest / list of point indices that the algorithm should operate on, rather than the entire cloud, via setIndices.

All classes inheriting from PCLBase exhbit the following behavior: in case no set of indices is given by the user, a fake one is created once and used for the duration of the algorithm. This means that we could easily change the implementation code above to operate on a <cloud, indices> tuple, which has the added advantage that if the user does pass a set of indices, only those will be used, and if not, the entire cloud will be used.

The new bilateral.hpp class thus becomes:

 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 #include #include ... template void pcl::BilateralFilter::applyFilter (PointCloud &output) { if (sigma_s_ == 0) { PCL_ERROR ("[pcl::BilateralFilter::applyFilter] Need a sigma_s value given before continuing.\n"); return; } if (!tree_) { if (input_->isOrganized ()) tree_.reset (new pcl::OrganizedNeighbor ()); else tree_.reset (new pcl::KdTreeFLANN (false)); } tree_->setInputCloud (input_); ...

The implementation file header thus becomes:

 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 #ifndef PCL_FILTERS_BILATERAL_IMPL_H_ #define PCL_FILTERS_BILATERAL_IMPL_H_ #include #include #include template double pcl::BilateralFilter::computePointWeight (const int pid, const std::vector &indices, const std::vector &distances) { double BF = 0, W = 0; // For each neighbor for (size_t n_id = 0; n_id < indices.size (); ++n_id) { double id = indices[n_id]; double dist = std::sqrt (distances[n_id]); double intensity_dist = abs (input_->points[pid].intensity - input_->points[id].intensity); double weight = kernel (dist, sigma_s_) * kernel (intensity_dist, sigma_r_); BF += weight * input_->points[id].intensity; W += weight; } return (BF / W); } template void pcl::BilateralFilter::applyFilter (PointCloud &output) { if (sigma_s_ == 0) { PCL_ERROR ("[pcl::BilateralFilter::applyFilter] Need a sigma_s value given before continuing.\n"); return; } if (!tree_) { if (input_->isOrganized ()) tree_.reset (new pcl::OrganizedNeighbor ()); else tree_.reset (new pcl::KdTreeFLANN (false)); } tree_->setInputCloud (input_); std::vector k_indices; std::vector k_distances; output = *input_; for (size_t i = 0; i < indices_->size (); ++i) { tree_->radiusSearch ((*indices_)[i], sigma_s_ * 2, k_indices, k_distances); output.points[(*indices_)[i]].intensity = computePointWeight ((*indices_)[i], k_indices, k_distances); } } #define PCL_INSTANTIATE_BilateralFilter(T) template class PCL_EXPORTS pcl::BilateralFilter; #endif // PCL_FILTERS_BILATERAL_H_

To make indices_ work without typing the full construct, we need to add a new line to bilateral.h that specifies the class where indices_ is declared:

 1 2 3 4 5 6 7 8 9 ... template class BilateralFilter : public Filter { using Filter::input_; using Filter::indices_; public: BilateralFilter () : sigma_s_ (0), ...

It is advised that each file contains a license that describes the author of the code. This is very useful for our users that need to understand what sort of restrictions are they bound to when using the code. PCL is 100% BSD licensed, and we insert the corpus of the license as a C++ comment in the file, as follows:

 1 * Copyright (c) XXX, respective authors.

## Proper naming

We wrote the tutorial so far by using silly named setters and getters in our example, like setSigmaS or setSigmaR. In reality, we would like to use a better naming scheme, that actually represents what the parameter is doing. In a final version of the code we could therefore rename the setters and getters to set/getHalfSize and set/getStdDev or something similar.

PCL is trying to maintain a high standard with respect to user and API documentation. This sort of Doxygen documentation has been stripped from the examples shown above. In reality, we would have had the bilateral.h header class look like: