Build a semantic segmentation dataset

Unofficial preview. Docsbook assembled this page from the public wkentaro/labelme README.md and examples/ directory. It is not affiliated with, endorsed by, or maintained by the Labelme project. The official documentation lives at labelme.io/docs.

Semantic segmentation labels every pixel with a class and does not distinguish one object from another of the same class. Two overlapping cars are simply car pixels. This is the examples/semantic_segmentation workflow.

Annotate the directory#

labelme data_annotated --labels labels.txt --validate-label exact \
  --config '{shape_color: {mode: auto, auto: {shift: -2}}}'

--validate-label exact rejects any label outside labels.txt, which matters more here than anywhere else: a stray class name becomes a stray channel in every mask you generate afterwards. The inline --config shifts the automatic per-label colours so adjacent classes are easier to tell apart while drawing.

Export class masks#

./labelme2voc.py data_annotated data_dataset_voc --labels labels.txt --noobject

--noobject is what makes this semantic rather than instance segmentation: it skips the per-object outputs and generates class-level masks only.

Why the mask looks black#

The label files contain only very low values such as 0, 4, 14, because those are class indices rather than brightness levels. The value 255 marks the __ignore__ label, which is -1 in the npy file.

Render one properly with the tutorial helper:

../tutorial/draw_label_png.py data_dataset_voc/SegmentationClass/2011_000003.png

Related#

• Instance segmentation — the same annotations, separated per object
• Export to Pascal VOC — every output directory the script can write
• Core concepts — why group_id decides which of the two tasks you get