RCI Image Contest
The goal of the RCI Image Contest is to compare the performance of current algorithms/pipelines for flagging, calibration, imaging and deconvolution in an equitable way on an even pre-agreed playing field. The goal is not to provide a battleground where one algorithm conquers all, but rather determine when and where certain algorithms can have an advantage, both now and in the future. Results for both image fidelity (using carefully defined metrics) and computational cost/speed will be assembled and published.
Steering Committee: Oleg Smirnov, Sanjay Bhatnagar, Torsten Enßlin, Urvashi Rao, Yves Wiaux, Gregg Hallinan, Katie Bouman, Vikram Ravi, Casey Law, Steven Myers
Project Scientist: Joshua Albert
Process
The RCI team will generate forward modeled data for a range of "true sky" images for the DSA-2000, ngVLA and SKA-mid. RFI (e.g. cell towers at a certain distance, satellites, etc.) will be injected and direction independent (electronics, baseline errors) and direction dependent (ionosphere, troposphere, antenna beam variations, pointing offsets) gain errors added. Sampling of the true sky images for the VLA Sky Survey, NVSS and other surveys (e.g ASKAP, MeerKAT) will also be generated to serve as initial models for calibration. Samples of these data will be made available to the community to test their algorithms. They will not have access to the true sky model! Participants can test a subset of the end-end pipeline (e.g. flagging only) or the full end-to-end pipeline and then submit a containerized version of their software to run on a reference machine to produce the final results. The Radio Interferometry Kaggle (RIK) will serve as the primary mechanism to make the RCI possible and to fund participation in the contest.
Examples of forward modeled data for 0.25% of the field of view of the DSA-2000. The sky model from Bonaldi et al. (2020; SKA Data Challenge) serves as the initial "True Sky" with the image centered on the brightest source in the ~7 square degree image. The observation simulated is for a single 15-minute integration, which serves as the basis for all DSA-2000 survey observations. The resulting thermal noise is 2 μJy. The bright unresolved source has a flux density of 380 mJy, thereby requiring dynamic range of 2 ⨉ 10⁵. The image shows the results after image-plane CLEAN for residual direction-independent phase errors of 0.1°, 1° and 10° (left to right)