Class labels¶

Models predict integer ids. These lists turn them back into names.

from kerasformers.utils.labels_util import COCO_80_CLASSES, ADE20K_150_CLASSES

print(len(COCO_80_CLASSES), COCO_80_CLASSES[0])
print(len(ADE20K_150_CLASSES), ADE20K_150_CLASSES[0])

80 person
150 wall

They live in kerasformers.utils.labels_util and are imported by path, not re-exported from kerasformers.utils. All are plain tuples, so they are cheap to import and safe to share.

Constant	Size	Used by
`COCO_91_CLASSES`	91	DETR, the DETR-family detectors, which keep COCO's original sparse ids
`COCO_80_CLASSES`	80	Detectors trained on the contiguous 80-class remap
`COCO_PANOPTIC_133_CLASSES`	133	Mask2Former, OneFormer, EoMT panoptic variants
`COCO_PANOPTIC_THING_IDS` / `COCO_PANOPTIC_STUFF_IDS`	80 / 53	Splitting a panoptic result into countable objects and background stuff
`ADE20K_150_CLASSES`	150	SegFormer, MaskFormer, the ADE20K semantic variants
`CITYSCAPES_19_CLASSES`	19	The Cityscapes-trained segmentation variants
`PASCAL_VOC_CLASSES`	21	DeepLabV3 and other VOC variants, index 0 being background

Usage¶

Most of the time you pass one straight into the post-processor rather than indexing it yourself, and read the names back out:

result = processor.post_process_object_detection(
    outputs, threshold=0.9, target_sizes=[(h, w)], label_names=COCO_80_CLASSES
)[0]
print(result["label_names"])

The detection post-processor defaults to COCO_91_CLASSES, so a model with a 91-class head needs nothing at all. Segmentation takes the list the same way, and plot_segmentation accepts one as class_names for its legend.

The 91 vs 80 trap¶

The two COCO lists are the single most common source of confidently wrong labels, because both produce plausible names for the same id. Original COCO ids run to 90 with gaps where categories were dropped; COCO_91_CLASSES preserves those gaps as "N/A" entries, 11 of them, while COCO_80_CLASSES closes them up.

print(COCO_91_CLASSES[11:14])
print(COCO_80_CLASSES[11:14])

('fire hydrant', 'N/A', 'stop sign')
('stop sign', 'parking meter', 'bench')

Index 0 is another giveaway: "N/A" in the 91-class list (ids start at 1), "person" in the 80-class one. The offset starts at zero and grows with each gap passed, so early classes look correct and later ones are quietly shifted, which is exactly the failure that survives a quick eyeball check. Match the list to the head's output width before trusting any of it.

Panoptic¶

COCO_PANOPTIC_133_CLASSES carries its group in the string, which keeps the legend readable when things and stuff sit side by side:

print(COCO_PANOPTIC_133_CLASSES[0])
print(COCO_PANOPTIC_133_CLASSES[80])

things: person
stuff: banner

The first 80 entries are things, the remaining 53 are stuff, and the two id tuples give you that split without parsing strings:

from kerasformers.utils.labels_util import (
    COCO_PANOPTIC_THING_IDS, COCO_PANOPTIC_STUFF_IDS,
)

print(COCO_PANOPTIC_THING_IDS[:5])
print(COCO_PANOPTIC_STUFF_IDS[:5])

(0, 1, 2, 3, 4)
(80, 81, 82, 83, 84)

They do not overlap and together cover all 133 ids. Use them to count instances (things only) or to build a background mask (stuff only) from a panoptic result.

See also Visualization utils and Utilities.