Learning-Deep-Learning

MapTR: Structured Modeling and Learning for Online Vectorized HD Map Construction

September 2022

tl;dr: Much improved version of VectorMapNet by introducing permutation-invariant loss.

Overall impression

VectorMapNet proposes to use polyline primitives (to be exact, directed polylines) to model map elements, and MapTR takes this idea further by also introducing permutation-invariant loss in the training.

Two different ways to handle set, which is inherently orderless. Either to process it with symmetric operation, such as maxpool or average-pool or min/max as in PointNet, or use min/max in the loss formulation to iterate through all permutations. –> The latter idea seems to have been popularized by monodepth2 when people realize they could plug in min/max operator in the loss formulation while keeping differentiability.

Instead of the coarse to fine two-stage method in VectorMapNet, MapTR uses instance queries and point queries and concatenates them to indicate a particular point from a particular instance. –> Maybe point queries is unnecessary, when polyline is viewed as a more complicated bbox.

Key ideas

Technical details

Notes