Hendrik

Masked Loss

6

Hey, Thanks for the great work in this! I have a question: I have GT segmentations that are wrong in some scans, and I know roughly where, but it is...

**Is your feature request related to a problem? Please describe.** A matcher class that matches like the naive matcher (so each prediction instance to the best reference) but then repeats...

Object Detection

1

Add support for bounding boxes - BBOX class with multiple instantiors - Transfer everything to one default internal representation (pixel-wise, not relative!) - Internally track what the data originally was...

**Is your feature request related to a problem? Please describe.** Currently, we have to define each metric in the PanopticaResult object as well. **Describe the solution you'd like** A way...

- Allow matching to yield all computed metrics to the instance evaluator - Allow matching to give matches based on multiple thresholds (so instance evaluator can immediately calculate AUTC values)...

Currently, we do not support arguments for Metrics. And we distinguish as evaluator arguments into global and instance metrics How about we overhaul the system to make it smoother: -...

Per sample AUTC values

1

- Calculate the matching value in matching step, dont give final matching - At the end, calculate the metrics for each threshold in range [0.0, 1.0] with default step-size 0.1...

**Describe the solution you'd like** Make a panoptica evaluator class for each input type, including instance proposals to easier get input-type specific behaviors (i.e. for semantic inputs, we do not...

Support for instance proposals (bounding boxes)

1

**Describe the solution you'd like** Add a datatype called boundingBoxPair, which will be converted to either semanticPair or unmatchedinstance pair internally after initialization. Additionally, we need metrics like AP, mAP,...