correspondence between .png annotation & category_id

[iTomxy]

It's said in Caffe-compatible stuff-thing maps that thing+stuff labels cover indices 0-181 and 255 indicates the 'unlabeled' or void class.
However, the mapping in labels.txt seems not match.

[iTomxy]

a simple demo:

import os
import sys
import numpy as np
import cv2
import matplotlib.pyplot as plt

P = "G:/dataset/COCO-stuff"
ANNO_P = os.path.join(P, "annotations/val2017")
LABEL_F = os.path.join(P, "cocostuff", "labels.txt")

id_cls = {}
with open(LABEL_F, "r") as f:
    for line in f:
        cid, cname = line.strip().split(": ")
        id_cls[int(cid)] = cname
        # print(cid, ":", cname)

mask_f = os.path.join(ANNO_P, "000000397133.png")
mask = cv2.imread(mask_f)
print(mask.min(), mask.max())
for c in range(mask.max()):
    m = (mask == c)
    area = m.sum()
    if area > 0:
        print(c, id_cls[c])  # `0` should be `person`, but output `unlabeled` !
        plt.imshow(np.where(m, c, 255))
        plt.show()
    break

[nightrome]

Yes, that is correct:

• Default dataset: 1-182 are valid classes.
• Caffe: 0-181 are valid classes.
• Separate stuff and thing downloads: Same as Default for stuff classes,, but the stuff annotations contain a class 'other' with index 183 that covers all non-stuff pixels..