bayer_demosaic operator incorrectly expects RGB/RGBA input image

[jaydcarlson]

In holoscan SDK 2.1.0, code was added in to ensure the input image has rank 3:

auto in_rank = in_tensor_gxf.rank();
if (in_rank != 3) {
    throw std::runtime_error(
       fmt::format("Input tensor has {} dimensions. Expected a tensor with 3 dimensions "
                   "(corresponding to an RGB or RGBA image).",
                   in_rank));
}

But this makes no sense. The input to a demosaic operator is a single-channel bayer image (rank 2).

Consequently, this check breaks the entirety of Holoscan Sensor Bridge, which relies on the demosaic operator in every example.

[tbirdso]

Hi @jaydcarlson , thank you for reporting this issue. Could you please confirm the version of Holoscan Sensor Bridge that you are using for reference?

From discussion with @AndreasHeumann , the rank check here is consistent with other tensor handling in Holoscan such as in Holoviz. Please consider the following workarounds:

• Use Holoscan Sensor Bridge 1.1-EA, which is compatible with Holoscan SDK v2.1.0
• Downgrade to Holoscan SDK ~v2.0.0~ v1.0.3 and use Holoscan Sensor Bridge 1.0

Update: Corrected Holoscan SDK version compatibility for Holoscan Sensor Bridge 1.0.

[grlee77]

Thank @jaydcarlson. I agree that it is confusing that a 3rd dimension (which I think must be size 1) is required. I can open an issue to update the operator to also accept 2D input in future releases (and still allow 3D as long the channels dimension is size 1)

[tbirdso]

Addressed in Holoscan SDK v2.4.0.