Approach to SpaceNet 6 challenge on instance segmentation. The pipeline follows Open Cities AI Challenge: Segmenting Buildings for Disaster Resilience.
15nd place out of 94 on public with 38.70937 jaccard index (top 1 -- 46.5162). Organizers decided to check on private test only top 10 participants on public. Here is an announcing the winners.
- GPU(s) with 16Gb RAM
- NVIDIA apex
The submission format satisfies the requirements as is in the submission template.
- Unet-like architecture with heavy encoder
efficientnet-b7
- Train on 512x512 crops with 4-channel SAR input
- 5 random folds (both train and test sets share the same Rotterdam city)
- Multi-channel masks: borders and contacts
- Binary-cross-entropy loss
- Predictions are made as
mask * (1 - contact) > 0.45
. It boosted score by 2 points over simplemask > 0.5