Starter kit for MLHEP-18 challenge
Repository contains starter kit for Semantic Segmentation of LArTPC tracks competition
| Phase № | Competition Type | Problem Type | Description | Data |
|---|---|---|---|---|
| 1 | Public | 3-class 3D semantic segmentation | You have events, each event is (192,192,192) 3D-tensor. Each cell in event can have signal and could be classified as 0 (no signal), 1 (electron/positron), 2 (all other particles) | train_1-2.hdf5 (6k events), test_1-2.hdf5 (4k events) |
| 2 | Private | 3-class 3D semantic segmentation | Private phase of competition. Closed for submissions. Submission from previous will be automatically migrated to this phase and will be validated on private dataset | --- |
| 3 | Public | 4-class 3D semantic segmentation | You have events, each event is (192,192,192) 3D-tensor. Each cell in event can have signal and could be classified as 0 (no signal), 1 (electron/positron), 3 (proton), 2 (all other particles) | train_3-4.hdf5 (10k events), test_3-4.hdf5 (10k events) |
| 4 | Private | 4-class 3D semantic segmentation | Private phase of competition. Closed for submissions. Submission from previous will be automatically migrated to this phase and will be validated on private dataset | --- |
| 5 | Public | 4-class 3D semantic segmentation with gaps | The same problem as in phases 3 and 4, but data contains the gaps | Data is not available yet |
| 6 | Private | 4-class 3D semantic segmentation | Private phase of competition. Closed for submissions. Submission from previous will be automatically migrated to this phase and will be validated on private dataset | --- |
All phases are evaluated with the following metrics with priority descend:
- acc_at_80 - 80-percentile of classification accuraces, averaged across each event (one event is (192,192,192) tensor)
- acc_at_50 - 50-percentile (median) of classification accuraces, averaged across each event
- acc_mean - mean classification accuracy, each accuracy is averaged across each event (one event is (192,192,192) tensor)
All metrics implemented in metrics.py script
Baselines for phases 1 (Public) and 2 (Private) are placed in phase_1-2 directory:
Baselines for phases 3 (Public) and 4 (Private) are placed in phase_3-4 directory:
pip install -r requirements.txt