- Task 1 - Single real-domain training (Models must be trained exclusively on the published Cityscapes dataset) - Method: Model selection
- Method info
method: Model selection2024-08-23
Authors: Michael Smith, Frank Ferrie
Affiliation: McGill University
Description: Use Hierarchical Multi-Scale Attention for Semantic Segmentation for the semantic segmentation only subsets, RbA (Mask2Former with RbA scoring) Swin-L for SMIYC, Ensemble of both (mean aggregation, softmax scoring) for synobjs. Takes advantage of the fact that BRAVO only requires confidence to be commensurate in each split. In a way this is essentially gaming the metric and competition structure.
This method is really intended as a comparison against more proper ensemble methods as to what is maximally possible given an ideal aggregation method (i.e. something better than the mean, as I do in entries Ensemble A and C).