[BUG] Default mask targets are not used when creating new segmentation annotation tasks

[geke-mir]

Describe the problem

The documentation states that datasets can have dataset.default_mask_targets or dataset.mask_targets here:

https://docs.voxel51.com/user_guide/using_datasets.html#storing-mask-targets

When creating a new segmentation task with FiftyOne without specifying mask_targets, classes 0-255 appear on CVAT. I doubt this default behavior is intuitive for users and perhaps this should work like annotating classifications, where users must specify classes or the annotation run crashes.

dataset.take(2).annotate("masks8", label_field="asdf", label_type="segmentation")

Code to reproduce issue

In [5]: from fiftyone import zoo

In [6]: dataset = zoo.load_zoo_dataset("quickstart")
Dataset already downloaded
Loading 'quickstart'
 100% |█████████████████████████████████████████████████████████████████████████████████████████████| 200/200 [1.6s elapsed, 0s remaining, 122.7 samples/s]         
Dataset 'quickstart' created

In [7]: dataset.default_mask_targets
Out[7]: {}

In [8]: dataset.take(2).annotate("jobbyjob", label_field="asdf", label_type="segmentation")

Computing metadata...
 100% |█████████████████████████████████████████████████████████████████████████████████████████████████| 2/2 [12.5ms elapsed, 0s remaining, 160.1 samples/s] 
Uploading samples to CVAT...
Out[8]: <fiftyone.utils.cvat.CVATAnnotationResults at 0x7e5936062090>

System information

• OS Platform and Distribution (e.g., Linux Ubuntu 22.04): Ubuntu 22.04.4 LTS
• Python version (python --version): 3.11.8
• FiftyOne version (fiftyone --version): 0.25.1
• FiftyOne installed from (pip or source): source

Other info/logs

The following diff seems to resolve the problem, but thought I'd open the issue first as this is more of a API/information flow change than a bug perhaps...

diff --git a/fiftyone/utils/annotations.py b/fiftyone/utils/annotations.py
index 9639da8d0..430af375a 100644
--- a/fiftyone/utils/annotations.py
+++ b/fiftyone/utils/annotations.py
@@ -844,10 +844,14 @@ def _parse_classes_dict(
 def _get_mask_targets(samples, mask_targets, label_field, label_info):
     if "mask_targets" in label_info:
         mask_targets = label_info["mask_targets"]
-
+    
+    # Check if we have defined a default mask target for the dataset
     if mask_targets is None:
-        mask_targets = {i: str(i) for i in range(1, 256)}
-        mask_targets[0] = "background"
+        if samples.default_mask_targets is not None:
+            mask_targets = samples.default_mask_targets
+        else:
+            mask_targets = {i: str(i) for i in range(1, 256)}
+            mask_targets[0] = "background"
 
     classes = [c for v, c in mask_targets.items() if v != 0]

Willingness to contribute

The FiftyOne Community encourages bug fix contributions. Would you or another
member of your organization be willing to contribute a fix for this bug to the
FiftyOne codebase?

• Yes. I can contribute a fix for this bug independently
• Yes. I would be willing to contribute a fix for this bug with guidance
  from the FiftyOne community
• No. I cannot contribute a bug fix at this time

[brimoor]

Hi @geke-mir 👋

I agree that it would be consistent with how the classes argument works if we did the following for the mask_targets argument:

When annotating new segmentation fields:

The user is required to provide mask_targets argument to annotate()

When annotating existing segmentation fields:

If the user doesn't provide mask_targets to annotate(), then:

• use mask targets from dataset.mask_targets[label_field] if they exist
• else fallback to dataset.default_mask_targets if those exist
• else fallback to the current default of:

mask_targets = {i: str(i) for i in range(1, 256)}
mask_targets[0] = "background"

A PR would be fantastic if you're open to contributing it! 😄

[geke-mir]

Thanks, Brian!

#4990