https://deeplearning.earth/series/oriented-det/Recent content in oriented-det on DeepLearning.EarthHugo -- gohugo.ioen-usSun, 06 Sep 2026 15:00:00 +0700https://deeplearning.earth/posts/2026-09-06_oriented_det_optical_satellite_demo/Sun, 06 Sep 2026 15:00:00 +0700https://deeplearning.earth/posts/2026-09-06_oriented_det_optical_satellite_demo/The oriented-det zoo now has four detector families. Three of them are worth putting on the same canvas: Rotated Faster R-CNN 3×, Rotated FCOS 3×, and Oriented R-CNN 3× — the accuracy leaders from v0.2.0. They are on a static page you can open in a browser. No Python, no GPU, no upload: dl4eo.com/object-detection-optical-satellite Pick a detector, pick a scene, then move the confidence slider. The boxes are precomputed; the page only filters what is drawn.https://deeplearning.earth/posts/2026-09-02_rotated_fcos_vs_oriented_rcnn_on_macos/Wed, 02 Sep 2026 15:00:00 +0700https://deeplearning.earth/posts/2026-09-02_rotated_fcos_vs_oriented_rcnn_on_macos/In June we ran Oriented R-CNN on a MacBook with --device mps — no CUDA toolchain, one CLI command, rotated boxes on a real aerial tile. v0.2.0 added a fourth detector family: Rotated FCOS, an anchor-free single-stage model with a decoded-rIoU 3× Hub checkpoint. This post puts both on the same Mac and the same images. Same canvas, same NMS, same score floor — Apple M1 Max, PyTorch MPS, Hub 3× weights.https://deeplearning.earth/posts/2026-08-28_oriented-det_v0_2_0_rotated_fcos_decoded_riou_and_the_updated_zoo/Fri, 28 Aug 2026 09:00:00 +0700https://deeplearning.earth/posts/2026-08-28_oriented-det_v0_2_0_rotated_fcos_decoded_riou_and_the_updated_zoo/Six weeks after v0.1.1, Oriented-Det v0.2.0 is on PyPI and tagged on GitHub. The headline is a new detector family: Rotated FCOS, an anchor-free single-stage model that sits in the zoo as the balanced pick — close to Rotated Faster R-CNN accuracy, without a region proposal network. This post is the release note: what landed, which Hub slug to download, and how the loss recipe (not the architecture name) is what moved the number.https://deeplearning.earth/posts/2026-07-11_oriented-det_v0_1_1_prob_iou_mmrotate_parity_and_the_updated_zoo/Sat, 11 Jul 2026 09:00:00 +0700https://deeplearning.earth/posts/2026-07-11_oriented-det_v0_1_1_prob_iou_mmrotate_parity_and_the_updated_zoo/Six weeks after v0.1.0, Oriented-Det v0.1.1 is on PyPI and tagged on GitHub. This is the release that packages the ProbIoU work, closes several MMRotate parity gaps in the training stack, and publishes the full eval-val protocol we have been using internally. If you already read Rotated Faster R-CNN on DOTA without custom CUDA, you have seen the technical story behind the headline number. This post is the release note: what changed, how to upgrade, and what to watch for.https://deeplearning.earth/posts/2026-07-10_rotated_faster_rcnn_probiou_dota/Fri, 10 Jul 2026 09:00:00 +0700https://deeplearning.earth/posts/2026-07-10_rotated_faster_rcnn_probiou_dota/Satellite scene with oriented bounding boxes Oriented object detection on satellite imagery lives or dies on rotated IoU — the overlap between two arbitrarily angled rectangles. Frameworks like MMRotate ship exact, CUDA-accelerated rotated IoU for training and inference. That is fast and precise, but it ties you to a heavy stack: MMCV custom ops, version pins, and compiled extensions that are painful to ship in a lean research codebase. OrientedDet takes a different path for v1: a full Python / PyTorch detector with no custom CUDA kernels and no MMCV runtime dependency.https://deeplearning.earth/posts/2026-06-29_announcing_the_final_oriented_det_pretrained_model/Mon, 29 Jun 2026 09:00:00 +0700https://deeplearning.earth/posts/2026-06-29_announcing_the_final_oriented_det_pretrained_model/We finished training and evaluation of Oriented R-CNN 3× on DOTA le90. With it, the OrientedDet pretrained zoo on Hugging Face (dl4eo/oriented-det-pretrained) is complete for our three detector families: Oriented R-CNN, Rotated Faster R-CNN, and Rotated RetinaNet. This post focuses on the new checkpoint — why it matters, how it compares to our 1× run and to MMRotate, how to load it, and how odet image-demo runs sliding-window inference on images larger than 1024×1024.https://deeplearning.earth/posts/2026-06-25_zero-shot_ship_detection_on_a_copernicus_sentinel-2_tile_with_oriented_rcnn/Thu, 25 Jun 2026 09:00:00 +0700https://deeplearning.earth/posts/2026-06-25_zero-shot_ship_detection_on_a_copernicus_sentinel-2_tile_with_oriented_rcnn/How far can a DOTA-pretrained oriented detector go on real Sentinel-2 imagery without any fine-tuning? We took a single 10 m true-colour crop from Copernicus Sentinel-2 — a busy patch of open water with ships at many headings — and ran it through the Oriented R-CNN model shipped with oriented-det. No maritime labels. No changes to the weights. Just a public checkpoint, a satellite tile, and a few inference knobs.https://deeplearning.earth/posts/2026-06-25_oriented_object_detection_on_macos_in_pure_python/Wed, 24 Jun 2026 09:00:00 +0700https://deeplearning.earth/posts/2026-06-25_oriented_object_detection_on_macos_in_pure_python/If you’ve ever tried to detect ships in a harbor or planes on a runway from an aerial photo, you’ve probably hit the same wall: axis-aligned boxes are a poor fit. Objects in satellite and drone imagery sit at arbitrary angles. You want rotated boxes — oriented bounding boxes — not rectangles forced to align with the image edges. For years, the go-to stack for this was MMRotate: powerful, but heavy.https://deeplearning.earth/posts/2026-06-23_oriented_rcnn_detections_for_the_15_dota_classes/Tue, 23 Jun 2026 09:00:00 +0700https://deeplearning.earth/posts/2026-06-23_oriented_rcnn_detections_for_the_15_dota_classes/Oriented object detection becomes concrete when you look at the boxes. In this post we will walk through detections produced by Oriented R-CNN on the 15 original classes from DOTA v1.0, the benchmark dataset that shaped much of the modern work on rotated object detection in aerial imagery. DOTA matters because it is not a neat toy dataset. Its images contain large scenes, dense object layouts, arbitrary object directions, and strong scale variation.https://deeplearning.earth/posts/2026-06-22_oriented-det_v0_1_0_sovereign_oriented_object_detection_for_eo/Mon, 22 Jun 2026 09:00:00 +0700https://deeplearning.earth/posts/2026-06-22_oriented-det_v0_1_0_sovereign_oriented_object_detection_for_eo/In May I announced that Oriented-det was coming. Today it is here. Oriented-Det v0.1.0 is officially released under the Apache 2.0 license, available on PyPI as oriented-det, and fully documented. This post covers the main ideas behind the framework and how to get started. What is Oriented-det? Link to heading Oriented-det is a lightweight PyTorch library for rotated object detection in aerial and satellite imagery. It targets the workflows where orientation actually matters — ships in a harbor, aircraft on an apron, vehicles in a dense parking lot — and where axis-aligned boxes simply miss the point.https://deeplearning.earth/posts/2026-05-28_introducing_oriented-det_sovereign_oriented_object_detection_for_eo/Thu, 28 May 2026 09:00:00 +0700https://deeplearning.earth/posts/2026-05-28_introducing_oriented-det_sovereign_oriented_object_detection_for_eo/Oriented-det is coming. Link to heading Oriented-det is a new offering for teams who need oriented object detection in Earth Observation (ships, aircrafts, vehicles) with a strong focus on sovereignty, license clarity, and time‑to‑deployment. I’m planning an official release for June 2026. Key selling points Link to heading Sovereign by design Designed to run where you need it: on‑prem, private cloud, regulated environments No “hosted inference” requirement and no platform lock‑in assumptions Open-source with a pragmatic license (Apache 2.