Error in `bounding_box.converts._xyxy_to_rel_xyxy`:

[LukeWood]

Currently this breaks when operating on ragged images. This can be see in the examples/layers/random_shear_demo.py demo.

More info available here: https://github.com/keras-team/keras-cv/pull/1537#issuecomment-1479032369

Perhaps we can unit test this behavior for all of our converters.

[james77777778]

Hi @LukeWood , @sebastian-sz

I think there might be a bug in the unit test: https://github.com/keras-team/keras-cv/blob/66d658ab214eaab9655026688a363c2b4052f916/keras_cv/bounding_box/converters_test.py#L108-L116

In this unit test, the behavior of bounding_box.convert_format is unusual. The source_box and target_box have a shape of (1, 2, 4), while the ragged_images have a shape of (2, (100, 50), (100, 50), 3).

_raggify should be applied to source_box and target_box before calling bounding_box.convert_format(), as shown below:

    @parameterized.named_parameters(*test_image_ragged)
    def test_converters_ragged_images(self, source, target):
        source_box = _raggify(boxes_ragged_images[source])
        target_box = _raggify(boxes_ragged_images[target])
        self.assertAllClose(
            bounding_box.convert_format(
                source_box, source=source, target=target, images=ragged_images
            ),
            target_box,
        )

After that, we can fix _image_shape as following:

def _image_shape(images, image_shape, boxes):
    if images is None and image_shape is None:
        raise RequiresImagesException()

    if image_shape is None:
        if not isinstance(images, tf.RaggedTensor):
            image_shape = tf.shape(images)
            height, width = image_shape[1], image_shape[2]
        else:
            height = tf.reshape(images.row_lengths(), (-1, 1))
            width = tf.reshape(
                tf.reduce_max(images.row_lengths(axis=2), 1), (-1, 1)
            )
            # remove the check if isinstance(boxes, tf.RaggedTensor)
            height = tf.expand_dims(height, axis=-1)
            width = tf.expand_dims(width, axis=-1)
    else:
        height, width = image_shape[0], image_shape[1]
    return tf.cast(height, boxes.dtype), tf.cast(width, boxes.dtype)

All unit tests have passed in my local env (pytest keras_cv/).

I can open a PR for this.

[LukeWood]

Ahhh - I think you are right. That makes sense how the test failed to catch the issue.