Semantic segmentation

  • Evaluation of Real-World Datasets:

We are not using all labels for comparing different algorithms since not all class labels exist in our real-world/synthetic datasets when these experiments were conducted. The following class labels were used: ground, building, tree, car, light pole, fence. Note that we are providing the implementation of KpConv on our Github to produce the baseline result shown below.

WMSC testing setup:

  1. Train in the real-world data sets except for WMSC. 

  2. Train in synthetic datasets (V1-V3). 

  3. Train in both real and synthetic datasets.

Evaluation of Real-World Datasets.JPG
  • Evaluation of Synthetic Datasets:

In light of the data diversity and quality, we selected the SyntheticV3 dataset for evaluation.
Training sets: Synthetic V3 (all - testing sets)

Testing sets: Synthetic V3 (5, 10, 15, 20, 25)

Evaluation of Synthetic Datasets.JPG

Instance segmentation

  • Evaluation of Synthetic V3:

We are providing the implementation of HAIS on our Github to produce the baseline result shown below. Please also check our instance segmentation challenge (Codalab) to participate and win the cash award!

Training sets: Synthetic V3 (all - testing sets)

Testing sets: Synthetic V3 (5, 10, 15, 20, 25)

Instance segmentation.JPG