[pcl::2d] Keypoint module implementation is broken and incompatible with updated Convolution/Edge/Kernel classes

[CelestialMelody]

Description

I encountered severe compilation errors and design inconsistencies when trying to use the pcl::2d::Keypoint module. It appears that the implementation of Keypoint in pcl/2d/impl/keypoint.hpp has not been updated to match the current API of its dependencies (pcl::2d::Convolution, pcl::2d::Edge, and pcl::2d::Kernel).

The code seems to rely on legacy methods that no longer exist or have been renamed, and there is a fundamental type incompatibility between Keypoint and Convolution.

Steps to Reproduce

1. Include pcl/2d/keypoint.h

2. Try to instantiate pcl::Keypoint with a standard PointCloud type or a vector type.

   #include <pcl/2d/keypoint.h>
   #include <pcl/point_cloud.h>
   #include <pcl/point_types.h>
   
   int main() {
     // Define the ImageType that meets the requirements of the source code (no longer PointCloud)
     using ImageType = std::vector<std::vector<float>>;
   
     int width = 20;
     int height = 20;
     ImageType input(height, std::vector<float>(width, 0.0f));
     ImageType output;
     
     for (int r = 5; r <= 15; ++r) {
       for (int c = 5; c <= 15; ++c) {
         input[r][c] = 255.0f;
       }
     }
   
     pcl::Keypoint<ImageType> keypoint;
     // Args: output, input, sigma_d, sigma_i, alpha, thresh
     keypoint.harrisCorner (output, input, 1.0f, 2.0f, 0.04f, 1e-5f); // compile error
   
     return 0;
   }

Detailed Errors

I have identified the following critical issues in pcl/2d/impl/keypoint.hpp:

1. Missing Methods in Convolution

The Keypoint implementation calls methods on the conv_2d object (type pcl::Convolution) that do not exist:

• Called: conv_2d.gaussianKernel(5, sigma, kernel_y);
  • Reality: pcl::Convolution has no gaussianKernel method. This method seems to belong to the pcl::kernel class.
• Called: conv_2d.convolve(output, input, kernel_y);
  • Reality: pcl::Convolution uses the standard PCL Filter interface. The method is likely filter(), or the class usage is completely wrong here.

2. Missing Methods in Edge

The code calls edge_detection.ComputeDerivativeXCentral(...). However, this method only has a declaration and has not been implemented.

3. Fundamental Type Incompatibility

The pcl::Keypoint class seems to be written assuming ImageType is a std::vector<std::vector<T>> (nested vector), as seen by the access pattern input[i][j] in keypoint.hpp.

However, pcl::Convolution and pcl::Edge explicitly require pcl::PointCloud<PointT> as their template argument and input types.

• If I pass pcl::PointCloud to Keypoint: keypoint.hpp fails because PointCloud does not support [i][j] indexing (it uses (col, row) or flat indexing).
• If I pass std::vector<std::vector<T>> to Keypoint: convolution.h fails because it expects a class with .width, .height and points members.

Environment

• OS: 6.18.5-arch1-1
• Compiler: gcc 15.2.1
• PCL: Master branch