1st Workshop on Scene Understanding
in Unstructured Environments
In conjunction with GCPR 2021, September 28 - October 1, 2021
Outdoor Semantic Segmentation Challenge
As part of the first workshop on scene understanding in unstructured environments we introduce the outdoor semantic segmentation challenge. The task is to develop an algorithm for a fine-grained vegetation and surface detection in an unstructured outdoor environment. The training data is provided by the TAS500 dataset. Participants are invited to send their predictions on the test set. All submissions are added onto our public leaderboard.
The best submissions will be invited so present their solution as part of the workshop at the DAGM GCPR 2021.
The submitted predictions are going to be evaluated on the TAS500 test split. For evaluation we will use the mean Intserection over Union (mIoU) metric. In addition, we will also display the novel Boundarry Jaccard (BJ) metric for each submission. The mean Boundary Jaccard (mBJ) focuses on evaluating pixels along class boundaries. This can be more relevant for fine-grained semantic segmentation tasks such as the scenes shown in the TAS500 dataset. Both the mIoU and the mBJ are highly correlated metrics. The final ranking on the leaderboard is done using the mIoU metric.
The Outdoor Semantic Segmentation Challenge is hosted on Codalab. CodaLab is an open-source web-based platform to collaborate on challenging resarch datasets. The public leaderboard and the automatic submission evaluation for this workshop challenge is hosted there.
We use Codalab's competition forum for discussions regarding the competition.
The challenge uses version 1.1 of the TAS500 dataset. The TAS500v1.1 dataset is available for download here.
More information on the TAS500 dataset is available on the project page.
All submissions are handled in the Codalab Competition. For more information visit the Participate section on Codalab.