Open michellerybak opened 3 months ago
There are some basic approaches. For example, you can feed an original image and a horizontally flipped image to the model, and measure the disagreement of model predictions. This disagreement can serve as confidence. You can also measure the disagreement between different models to probe the difficulty of the sample. You may refer to the research area of uncertainty estimation for regression tasks.
Is there a way to estimate a confidence associate with a pixel-level depth prediction?