Rebeca Villarraso

Computer Vision Testing Analyst

1 minute read

Objectives

  1. Continue adapting DeepLabV3+ with the RELLIS-3D dataset.
  2. Update Github Pages.

Progress This Week

1. DeepLabV3+ on Rellis-3D

This week I was able to adapt Rellis-3D to the DeepLab V3 network. The problems I have been able to solve are related to the source labels of the data set, the network expects labels numbered from 0-19 and the labels of the data set have different values.

To solve this problem, a script has been created that maps the identifiers of the original labels to new ones from 0-19 and generating the new mask images. The new masks with their corresponding identifiers have been used as input to the DeepLabV3+ network and the training has been carried out with few epochs successfully.

2. Github Pages

Updated post in Github Pages.

3. Goose Dataset

This week I was also able to download the Goose dataset (which is now available) and prepare it to perform semantic segmentation with DeepLabV3+.

Next Week Work Planning

  1. Add the training and test data on the same “Accuracy” graph.
  2. Add training and validation data on the same “Loss” graph.
  3. Add early stopping in train process.
  4. Add model checkpoints/callbacks in train process.
  5. Add Tensorboard to monitoring training process.
  6. Update GitHub Pages.

References

  • [Cityscapes Metrics] https://www.cityscapes-dataset.com/benchmarks/#instance-level-scene-labeling-task