fcclip/data/datasets/register_pascal_voc_21_semantic.py

# Copyright (c) Facebook, Inc. and its affiliates.
import os

import numpy as np

from detectron2.data import DatasetCatalog, MetadataCatalog
from detectron2.data.datasets import load_sem_seg

from . import openseg_classes

PASCAL_VOC_21_CATEGORIES = openseg_classes.get_pascal_21_categories_with_prompt_eng()

PASCAL_VOC_21_COLORS = [k["color"] for k in PASCAL_VOC_21_CATEGORIES]

MetadataCatalog.get("openvocab_pascal21_sem_seg_train").set(
    stuff_colors=PASCAL_VOC_21_COLORS[:],
)

MetadataCatalog.get("openvocab_pascal21_sem_seg_val").set(
    stuff_colors=PASCAL_VOC_21_COLORS[:],
)


def _get_pascal21_meta():
    # Id 0 is reserved for ignore_label, we change ignore_label for 0
    # to 255 in our pre-processing, so all ids are shifted by 1.
    stuff_ids = [k["id"] for k in PASCAL_VOC_21_CATEGORIES]
    assert len(stuff_ids) == 21, len(stuff_ids)

    # For semantic segmentation, this mapping maps from contiguous stuff id
    # (in [0, 91], used in models) to ids in the dataset (used for processing results)
    stuff_dataset_id_to_contiguous_id = {k: i for i, k in enumerate(stuff_ids)}
    stuff_classes = [k["name"] for k in PASCAL_VOC_21_CATEGORIES]

    ret = {
        "stuff_dataset_id_to_contiguous_id": stuff_dataset_id_to_contiguous_id,
        "stuff_classes": stuff_classes,
    }
    return ret


def register_all_pascal21(root):
    root = os.path.join(root, "pascal_voc_d2")
    meta = _get_pascal21_meta()
    for name, dirname in [("train", "training"), ("val", "validation")]:
        image_dir = os.path.join(root, "images", dirname)
        gt_dir = os.path.join(root, "annotations_pascal21", dirname)
        name = f"openvocab_pascal21_sem_seg_{name}"
        DatasetCatalog.register(
            name, lambda x=image_dir, y=gt_dir: load_sem_seg(y, x, gt_ext="png", image_ext="jpg")
        )
        MetadataCatalog.get(name).set(
            stuff_classes=meta["stuff_classes"][:],
            thing_dataset_id_to_contiguous_id={},  # to make Mask2Former happy
            stuff_dataset_id_to_contiguous_id=meta["stuff_dataset_id_to_contiguous_id"],
            image_root=image_dir,
            sem_seg_root=gt_dir,
            evaluator_type="sem_seg",
            ignore_label=255,
            gt_ext="png",
        )
        
_root = os.getenv("DETECTRON2_DATASETS", "datasets")
register_all_pascal21(_root)