Inconsistency between PyCOCOCallback and BoxCOCOMetrics

[tu6berk]

Current Behavior:

While constructing the ground truth dataset, PyCOCOCallback computes the number of detections within annotations as follows:

num_detections = ops.sum(ops.cast(gt_classes > 0, "int32"), axis=-1)

at line 102.

In constrast, BoxCOCOMetrics tackles it via this approach:

ground_truth["num_detections"] = [ ops.sum(ops.cast(y_true["classes"] >= 0, "int32"), axis=-1) ]

at lines 300-302.

Consequently, the results produced by the two are quite different, especially when your data set has only a few distinct labels and the encoding is plain integer assignment with enumeration starting from 0.

Expected Behavior:

Since label encoding schemes in most Keras CV models enumerate the classes starting from 0, which holds in my case as well, I believe that BoxCOCOMetrics approach should be applied to PyCOCOCallback. Alternatively this could be made a parameter to be determined by the user.

Steps To Reproduce:

Version:

• 0.8.2

Anything else:

[tu6berk]

hi, any updates, please?

[LukeWood]

I also agree that the PyCOCOCallback is incorrect here, and that it should be updated to match the BoxCOCOMetrics (Assuming that the package still uses 0 as the first class index across the repo)