Missed Detection in Ultralytics vs GitHub/HF weights

[reachpranjal]

I find some differences in inference results between Ultralytics release and GitHub weights. I'm referring to the "building" text prompt which the ultralytics version completely misses for all the versions (s, m, l, x).

Ultralytics Weights

GitHub Weights or Huggingface Weights

Have raised related issue in ultralytics here: #18331