Evaluation Metrics

1 Threshold accuracy 1 (${\delta }_1$)

$$\text{\% of}{d}_i,\text{s.t.}max(\frac{d_i}{d_i^{\ast }},\frac{d_i^{\ast }}{d_i})=\delta <1.25$$

where $d_i$ refers to the prediction depth of the pixel $i$ and $d_i^{\ast }$ refers to the ground truth, the same follow.

2 Threshold accuracy 2 (${\delta }_2$)

$$\text{\% of}{d}_i,\text{s.t.}max(\frac{d_i}{d_i^{\ast }},\frac{d_i^{\ast }}{d_i})=\delta <{1.25}^2$$

3 Threshold accuracy 3 (${\delta }_3$)

$$\text{\% of}{d}_i,\text{s.t.}max(\frac{d_i}{d_i^{\ast }},\frac{d_i^{\ast }}{d_i})=\delta <{1.25}^3$$

4 RMSE

RMSE can reflect the absolute error between predicted depths and ground truths,

$$RMSE=\sqrt{\frac{1}{\Vert \mathbb{N}\Vert }\sum _{i\in \mathbb{N}}{\Vert d_i-d_i^{\ast }\Vert }^2}$$

where $\mathbb{N}$ refers to the set of pixels in depth maps, the same follow.

5 RMSE Log

Using logarithmic operations on the RMSE to obtain the RMSE Log, the latter can reduce the effect of outliers when measuring prediction error.

$$RMSELog=\sqrt{\frac{1}{\Vert \mathbb{N}\Vert }\sum _{i\in \mathbb{N}}{\Vert \log d_i-\log d_i^{\ast }\Vert }^2}$$

6 Abs Rel

Abs Rel refers to the relative error between the predicted depth and the ground truth.

$$AbsRel=\sum _{i\in \mathbb{N}}\frac{|d_i-d_i^{\ast }|}{d_i^{\ast }}$$

7 Sq Rel

Using the squaring operation on Abs Rel yields Sq Rel, which reflects the prediction error more significantly.

$$SqRel=\sum _{i\in \mathbb{N}}\frac{{\Vert d_i-d_i^{\ast }\Vert }^2}{d_i^{\ast }}$$