[Docs] single-scale DOTA-v1.0 mAP

[2towardsnorth]

Branch

master branch https://mmrotate.readthedocs.io/en/latest/

📚 The doc issue

Could you please clarify the source of the 77.49% single-scale DOTA-v1.0 mAP reported for LSKNet-S in several later papers? The original LSKNet paper reports 81.64% with multi-scale training/testing, but does not appear to include the single-scale result. Was 77.49% obtained by the LSKNet authors, or was it reproduced by a later work? Is the corresponding config or log available?

Suggest a potential alternative/fix

No response

[zcablii]

Thanks for raising this question. The 77.49% single-scale DOTA-v1.0 mAP result was not reported or reproduced by the official LSKNet authors. It appears to originate from third-party implementations or subsequent works that evaluated LSKNet-S under their own single-scale settings.

For the exact experimental setup and reproducibility of the 77.49% result, we would recommend referring to the third-party work that first reported it or its associated implementation.

[2towardsnorth]

Thanks for raising this question. The 77.49% single-scale DOTA-v1.0 mAP result was not reported or reproduced by the official LSKNet authors. It appears to originate from third-party implementations or subsequent works that evaluated LSKNet-S under their own single-scale settings.

For the exact experimental setup and reproducibility of the 77.49% result, we would recommend referring to the third-party work that first reported it or its associated implementation.
I achieved a score of 78.49% on the test set using single-scale evaluation on an RTX 4090, but I couldn’t find any details about the third-party implementation.