Incorrect offset calculation in PointField when creating PointCloud2 messages

[tokuda99]

Description

While using the library to convert PCD data into ROS PointCloud2 messages, I encountered an issue where the offset values for the PointField fields were calculated incorrectly. This resulted in fields like intensity or other attributes being misaligned in the generated PointCloud2 message, causing visualization tools like RViz to display incorrect or missing data.

Expected Behavior

The offset for each PointField should reflect the cumulative byte position of the field within a point, considering the size and count of all preceding fields.

Actual Behavior

The current implementation calculates offset using i * type_.itemsize, which assumes a fixed-size field alignment without considering the actual data layout or field sizes. This leads to incorrect offset values for fields that do not align sequentially (e.g., fields with different sizes or counts).

Steps to Reproduce

1. Use the library to generate a PointCloud2 message from a PCD file with fields of varying sizes and counts, such as:

• x, y, z (float32)
• intensity (uint8)
• return_type (uint8)
• channel (uint16)

2. Publish the PointCloud2 message in ROS.
3. Attempt to visualize the data in RViz or CloudCompare.

Example Code

Here is the problematic snippet:

for i, (field, type_, count) in enumerate(zip(self.fields, self.types, self.counts)):
    type_ = np.dtype(type_)

    fields.append(
        PointField(
            name=field,
            offset=i * type_.itemsize,  # Incorrect offset calculation
            datatype=NPTYPE_TO_PFTYPE[type_],
            count=count,
        )
    )

Suggested Fix

The offset should be calculated as the cumulative sum of the sizes of all preceding fields, adjusted for the field's count. For example:

offset = 0

for field_name, type_str, count in zip(self.fields, self.types, self.counts):
    np_type = np.dtype(type_str)
    
    fields.append(
        PointField(
            name=field_name,
            offset=offset,  # Correct cumulative offset
            datatype=NPTYPE_TO_PFTYPE[np_type],
            count=count,
        )
    )
    
    offset += np_type.itemsize * count  # Accumulate size

Impact

This issue causes data misalignment in the PointCloud2 message, leading to:

• Visualization tools misinterpreting fields (e.g., intensity not showing up or being assigned wrong values).
• Potential bugs in downstream processing nodes that rely on correctly aligned fields.

Please let me know if you need further details or examples to reproduce the issue. Thank you for addressing this!

[urasakikeisuke]

@tokuda99
Thank you for sharing your findings. It looks like a solid solution.
Could you please open a pull request with your changes so we can review them together?