niqe/ilniqe returns values outside of range

[hdrmadness]

Hello

I am getting large niqe and ilniqe scores, so I put some print statements

def forward(self, X: torch.Tensor) -> torch.Tensor: r"""Computation of NIQE metric. Args: X: An input tensor. Shape :math:(N, C, H, W). Returns: Value of niqe metric in [0, 1] range. """ print(X.size()) score = calculate_niqe(X, self.crop_border, self.test_y_channel, self.pretrained_model_path, self.color_space) print("niqe score should be 0 to 1, and we got {}".format(score)) return score

and I get:

torch.Size([1, 3, 512, 344]) | 1/10 [00:01<00:15, 1.78s/it] niqe score should be 0 to 1, and we got 5.133647061808286

[chaofengc]

Sorry that the document is not correct. The output value of niqe/ilniqe is a positive real value, without explicit upper bound limitation.

[hdrmadness]

Thanks for looking into this. Is there a reference in your documentation where the bounds of each metric can be found? I am finding it difficult to correlate scores of different metrics to each other. I attached a list that I am looking at.

NR-IQA bounds.txt

[chaofengc]

Thanks for your information. I have added it to the ModelCard

There are several points about the metric bound:

• The absolute value of each metric, including its bound, is usually meaningless. Only the relative scores of different images are meaningful. The absolute value is related to the training data used as well as some math operator such as sigmoid.
• There are no standard settings in IQA currently. Therefore, we do not compare absolute values between different metrics.
• To correlate scores of different metrics to each other, we usually used the PLCC and SRCC which measures their linear correlations or ranking correlations.