Image size assumption causes memory access error in Linemod object recognition

On my Ubuntu 13.04 (64 bit) system, Linemod object recognition started to crash irregularly, after applying a region of interest on my input data.

Reproduction:

• Construct a pointcloud with a size of w * h where w%16 or h%16 does not equal 0.
• Create templates using Linemod
• Try to match the templates on the input data
• This bug appears seemingly at random

Valgrind output:

        ==25525== Invalid write of size 2
        ==25525==    at 0xA947A36: pcl::LINEMOD::matchTemplates(std::vector<pcl::QuantizableModality*, std::allocator<pcl::QuantizableModality*> > const&, std::vector<pcl::LINEMODDetection, std::allocator<pcl::LINEMODDetection> >&) const (linemod.cpp:288)
        ==25525==    by 0x4E7D782: Linemod::matchTemplates(boost::shared_ptr<pcl::PointCloud<pcl::PointXYZRGBA> const> const&) (linemod.cpp:166)

Solution:

In the Linemod object recognition (line 271), memory is read 16 bits at a time, causing memory errors when the input data's width or height are not a multiple of 16.

This is a hidden assumption and could either:

• be documented
• exception/error handling could be implemented, warning the user
• the best solution would be to check the input cloud size and correctly handle the last set of bits (with length < 16)

This issue actually appeared in OpenCV as well