https://deeplearning.earth/tags/hrsc2016/Recent content in Hrsc2016 on DeepLearning.EarthHugo -- gohugo.ioen-usSun, 13 Sep 2026 16:00:00 +0700https://deeplearning.earth/posts/2026-09-13_hrsc2016_recipes_trains_and_results/Sun, 13 Sep 2026 16:00:00 +0700https://deeplearning.earth/posts/2026-09-13_hrsc2016_recipes_trains_and_results/DOTA is the pretrain zoo. HRSC2016 is the small-data ship zoo. Three oriented-det families now have native HRSC recipes and published 3× Hub weights: Oriented R-CNN 90.41%, Rotated Faster R-CNN 88.77%, Rotated FCOS 88.34% mAP50. Those numbers are make eval-val on ImageSets test. Test is not in training. That sentence is the whole point of this dataset in our stack — DOTA eval-val is leaky; HRSC eval-val is a real holdout.