LabeledEuclideanClusterExtraction memory allocation

The LabeledEuclideanClusterExtraction demands a preallocalization of the output vector (vector of vector of point indices) in the number of labels (see this access ).

Although infeasible (or at least impracticable) for small labels, at least in the case of the label std::numeric_limits<uint32_t>::max() (which is a valid label) it becomes quite apparent that this won't work.

Your Environment

• Operating System and version: Arch Linux
• Compiler: clang 5.0
• PCL Version: 1.8.0

Expected Behavior

1. Work with minimum necessary preallocalizations.
2. Don't segfault without external preallocalizations.

Current Behavior

1. Infeasible memory / address space requirements.
2. Segfaults.

Possible Solution

The expected outcome is a mapping of labels to segments. There is a reason std::map exists...

Code to Reproduce

    extraction_t extr;
    extr.setClusterTolerance(irrelevant);
    extr.setMinClusterSize(1);
    extr.setInputCloud(cloud);
    std::vector<std::vector<pcl::PointIndices>> result;
    extr.extract(result);

Start at segmentation/include/pcl/segmentation/impl/extract_labeled_clusters.hpp around the linked access and reproduce the issue with the provided extraction example. Trace how labels determine the output vector allocation and verify behavior with a label of std::numeric_limits<uint32_t>::max(). Done means extraction avoids infeasible allocation and does not segfault without external preallocation.