Depth Completion Notes

The current evaluation metrics (RMSE and MAE) can be quite misleading. Mixed-depth or blurry depth map outputs may lead to lower RMSE, and are only weakly penalized in MAE (see Section 3 of Depth Coefficients for Depth Completion). Let's take a closer look at the point clouds created by re-projecting the depth maps into 3D space:

RMSE and MAE metrics are almost the same for the Typical Deep Network and Ours, but structure is clearly preserved in our method.

---

Orange: Pedestrians and cyclist (note the bicycle wheels) Green: Thin poles, Red: Even far away cyclists are reasonably recovered.

Advantages to our method:

• Fast, runs on CPU only
• Structure preserving
• No training data required
• Only LiDAR data required

Disadvantages:

• Does not use RGB or LiDAR intensity information, which can provide useful semantic information for better reconstruction
• Shape of some objects are not fully recovered when LiDAR points are lost

How to tell if a method is structure preserving on the test server:

• Test Image 1: Check for snowmen (Do the pedestrians look like pedestrians or large blobs?)
• Test Image 6: Closest pole should have very low D1 Error (Do the depths show a pole, or has it been warped?)