Semantic Segmentation of LArTPC tracks
Previous
Public 3Current
Private 3End
Competition EndsWhy segmenting pixels?
In the first step of this challenge we ask you to classify non-zero pixels into two basic category of particles: energy deposited by electron/positron, referred to as EM-particle, vs. all other particles. An accurate identification of EM-particle pixels is a crucial task to identify electron neutrino interaction for neutrino oscillation experiments using LArTPC detectors. In a traditional data reconstruction process of LArTPC experiments, this distinction is made after pixels are clustered into individual particles and analyzing the topological feature of clustered pixels. However, this is proven to be difficult. Instead, having a pixel-level distinction of EM-particles beforehand can improve the performance of clustering and simplify the rest of data reconstruction chain.
At the second step of the challenge, we will add another distinct label to those pixels that contain energy deposited by protons. Two most basic yet important neutrino interaction final states contain electron+proton from electron neutrino interaction, or muon+proton from muon neutrino interaction. Adding the proton label therefore improve the separation power between two interaction channels.
Finally, at the third step of the challenge, we will make the simulation sample more realistic by introducing gaps in the data sample which represents an unresponsive part of the detector. Your algorithm needs to overcome this lack of information in order to be proven useful toward an application to real data.
Evaluation
The problem is a 3D multiclass semantic segmentation problem. Each sample is 3D (192,192,192) tensor. You must predict the following classes for each cell ("pixel"): no signal (0), electron/positron (1) [, proton (3)], another particles (2).
There are 6 phases:
Phase 1: public phase. During this phase you need to build model for 3-class semantic segmentation: no signal (0), electron/positron (1) and another kind of particles (2). You are given two datasets: train.hdf5 ("data" and "label" columns) and test.hdf5 ("data" column only). You need to provide submission .hdf5 file with "pred" column, containing 3D (192,192,192) tensors with predicted labels (predictions at each cell must be in discrete range {0, 1, 2} !)
Phase 2: private phase. No manual submission needed - submission from previous phase will be automatically validated on private holdout. This is the final phase of first competition step (3-class segmentation) used for final leaderboard evaluation.
Phase 3: public phase. The data structure is the same as in previous step. The only difference is that now You have predict 4 classes: no signal (0), electron/positron (1), proton (3) and another kind of particles (2). Predictions in submission file must be in discrete ({0,1,2,3}) range!
Phase 4: private phase. Works in the same way as phase 2
Phase 5: public phase. The problem and data have the same structure as in "Phase 3", but now data contains the gaps (corresponds to unresponsive part of the detector)
Phase 6. private phase. Works in the same way as in phase 2 and 4
Rules
Submissions must be made before the end of each public phase (phase № 1, 3 and 5). You may submit 10 submissions every day and 100 in total for each phase.
Starter kit
Starter kit is available on github.
Public 1
Start: Aug. 6, 2018, midnight
Description: Public leaderboard (1st Cycle)
Private 1
Start: Aug. 9, 2018, 1 a.m.
Description: Private leaderboard (1st Cycle)
Public 2
Start: Aug. 9, 2018, 1 a.m.
Description: Public leaderboard (2nd Cycle)
Private 2
Start: Aug. 12, 2018, 1 a.m.
Description: Private leaderboard (2nd Cycle)
Public 3
Start: Aug. 12, 2018, 11 a.m.
Description: Public leaderboard (3rd Cycle)
Private 3
Start: Oct. 2, 2018, 1 a.m.
Description: Private leaderboard (3rd Cycle)
Competition Ends
Never
You must be logged in to participate in competitions.
Sign In