Surface Normal Estimation

---

1. Task

Given a single RGB image, recover the pixel-wise surface normals, and decide the relative relations among predicted normals, i.e. orthognal, parallel, or neither.

2. Evaluation Metrics

Absolute Normal Error:

• Mean: The mean of angular error measured in degrees

• Median: The mean of angular error measured in degrees

• 11.25°: the percentage of pixels whose angular error is within 11.25 degrees.

• 22.5°: the percentage of pixels whose angular error is within 22.5 degrees.

• 30°: the percentage of pixels whose angular error is within 30 degrees.

Relative Normal Error:

• AUC _o : Area Under Curve for the orthogonal normal pairs.

• AUC _p : Area Under Curve for the parallel normal pairs.

For a more detailed explanation of the above metrics, please refer to the OASIS paper.

3. Leaderboard

Method	Mean	Median	11.25°	22.5°	30°	AUC o	AUC p
Hourglass (OASIS)[1]	23.91	18.16	31.23%	59.45%	71.77%	0.5913	0.5786

References

^[1]

Hourglass (OASIS)

Chen, W., Qian, S., Fan, D., Kojima, N., Hamilton, M., & Deng, J. OASIS: A Large-Scale Dataset for Single Image 3D in the Wild. CVPR 2020.