List view
https://neurips.cc/Conferences/2021/Schedule?type=Workshop
Overdue by 4 year(s)•Due by October 2, 2021**1. Downstream task** 1.1 Get the seg code working * 1.2 Get the seg model to converge * 1.3 Train/test on the S.E labels that Lucas downloaded * 1.4 Train supervised model (from scratch). 1.5 All S2 bands 1.6 Cloud mask, L2A includes shadows in cloud masks. **2. Metrics, monitoring, reporting, post-mortem** **3. Embeddings of time series** 3.1 Preprocessing: consolidate time-series at a monthly basis (1 image/month) ** 3.2 Backbone for time-series: RNN, 3DCNN, transformer ** 3.3 Naive: average revisit-specific embeddings. 3.4 Consolidate cloud masks from L2A. 3.5 Invariance to cloud-coverage. **4. Fusion** 4.1 Weather data (ERA-5, temp, precipitation) 4.2 S1 (SAR) **5. Visualization** 5.1 End-user visualisation. 5.2 Re-use the ML4CC dashboard to viz the model output.
No due date- AOIs? - Dynamic Earth Net: http://www.classic.grss-ieee.org/earthvision2021/challenge.html - Ground truth data? - S2Sen - Embed the time-series, or time-series of embeddings? - What is our validation/testing pipeline for embeddings? - Inspiration: SeCo paper. - How do we deal with clouded / "missing" time-points? - Cloud-mask?
Overdue by 4 year(s)•Due by July 9, 2021