Wrong implementation of median frequency re-weighting.

[nnop]

TensorFlow-ENet/get_class_weights.py

Lines 88 to 94 in 8bb8322

	for i, j in label_to_frequency_dict.items():
	j = sorted(j) #To obtain the median, we got to sort the frequencies
	median_frequency = np.median(j) / sum(j)
	total_frequency = sum(j) / total_pixels
	median_frequency_balanced = median_frequency / total_frequency
	class_weights.append(median_frequency_balanced)

The implementation is not consistent with that stated in (Eigen and Fergus, 2015) :

Reference:

• D. Eigen and R. Fergus, “Predicting depth, surface normals and semantic labels with a common multi-scale convolutional architecture,” in ICCV, pp. 2650–2658, 2015.

[kwotsin]

Do you specifically refer to line 91 where the median is divided by sum(j)?

[nnop]

You should take the median of frequencies among all classes, not the median among all images of a single class. Meanwhile, not needing sorting before use np.median.