Problem Description
Large pixel masks, particularly those obscuring the stellar core, can bias centroid estimation and degrade the quality of the training sample. Since WaveDiff assumes reliable stellar observations during training, heavily obscured stars may adversely affect both model convergence and reconstruction accuracy.
An automated method is therefore required to identify and reject stars whose masked regions exceed acceptable limits.
Proposed Solution
Implement a configurable preprocessing step that evaluates the level of pixel mask obscuration for each star.
The feature should:
- compute one or more mask obscuration metrics (e.g., fraction of masked pixels within configurable apertures);
determine whether the stellar core is significantly affected;
- classify observations as acceptable or outliers using configurable thresholds;
- exclude rejected observations from downstream training (and optionally validation);
- generate summary statistics describing the number and fraction of rejected stars.
The implementation should remain modular so that additional outlier rejection criteria can be incorporated in future milestones.
Alternatives Considered
- Manual inspection — impractical for operational pipeline execution.
- Reject stars based only on the total masked fraction — simple but does not distinguish between peripheral masking and masking of the stellar core.
- Learn masking effects during model training — increases model complexity and does not prevent poor-quality observations from influencing optimisation.
Additional Context
This feature forms one component of the Outlier Rejection milestone.
The initial milestone focuses on two complementary rejection criteria:
- Pixel mask obscuration
- Poor goodness-of-fit using a previously trained WaveDiff model
Both methods should be configurable and designed to integrate into the preprocessing stage prior to model training.
Tasks to complete
Problem Description
Large pixel masks, particularly those obscuring the stellar core, can bias centroid estimation and degrade the quality of the training sample. Since WaveDiff assumes reliable stellar observations during training, heavily obscured stars may adversely affect both model convergence and reconstruction accuracy.
An automated method is therefore required to identify and reject stars whose masked regions exceed acceptable limits.
Proposed Solution
Implement a configurable preprocessing step that evaluates the level of pixel mask obscuration for each star.
The feature should:
determine whether the stellar core is significantly affected;
The implementation should remain modular so that additional outlier rejection criteria can be incorporated in future milestones.
Alternatives Considered
Additional Context
This feature forms one component of the Outlier Rejection milestone.
The initial milestone focuses on two complementary rejection criteria:
Both methods should be configurable and designed to integrate into the preprocessing stage prior to model training.
Tasks to complete