Wedding, soap bubbles

Where a bubble crosses the couple, two surfaces share one pixel. A depth map holds one number per pixel, so transparency has no correct answer - only a choice, and models make different ones.

Table of Contents

The picture
Where it shows
How the models read it
The other scenes

The picture

Compare models on this scene Every map of it at once

Where it shows

The difficulty is not spread evenly over the frame. Each link opens the comparison already zoomed to one of these places.

Bubble across the groom Bubble on the dress

How the 13 shipped models read it

Each at its finest setting. The differences are the point - and the reason there is no score under any of them: a generated image has no ground-truth depth to score against, so what a map gets right is a judgment, not a number.