RESEARCH / REPRODUCIBILITY
Evidence first. Claims second.
The current repository contains testable perception software and scientific research material. No rights-cleared maritime model run or independent acoustic evaluation is verified.
REAL FRAMES: 0 · OFFICIAL BENCHMARK: NOT RUNPerception evidence
| MEASURE | OBSERVED | INTERPRETATION |
|---|---|---|
| Rights-cleared real frames evaluated | 0 | No accuracy or robustness conclusion |
| mIoU / obstacle F1 | Not measured / Not measured | Requires approved data and official protocol |
| Named-hardware FPS | Not measured | End-to-end timing is not available |
| Local API integration tests | Synthetic fixtures only | Software behavior; not maritime performance |
Acoustics archive
The provided archive labels 96 patches across 20 source images sharing one capture-time token. Patch-random splits can leak correlated data. No independent recording-session split has been evaluated, and the H5 weight file has not been deserialized.
Selected upstream work
- eWaSR — candidate embedded baseline; no checkpoint included or loaded.
- WaSR and WaSR-T — published comparators; no results reproduced here.
- MaCVi datasets — access and use terms need per-dataset review.
The software snapshot comes from the historical private IRISEA repository. It includes integration code and experimental interpretation logic, but no trained proprietary segmentation weights or verified performance gain.