Which oriented detector should you train?

If I were starting a train tomorrow, I would start with Rotated RTMDet-S.

On official DOTA Task 1 it scores 77.60%. A box counts when it covers at least half the real object. That score is AP50. The tighter score, AP75, asks for three quarters of the object, and RTMDet reaches 52.01. Both lead this series. The checkpoint is rotated_rtmdet_dota_le90_3x. I show a box when its score is at least 0.45.

I switch only when the picture asks for it.

• A diagonal row of buses or trucks, where each box has to follow the vehicle → Oriented R-CNN, score 0.55.
• Finding the object matters more than a perfect box, on a ResNet model → Faster R-CNN, score 0.60.
• That same diagonal row, on a shorter ResNet train → FCOS, score 0.20.

The number behind each pick is below.

Start with Rotated RTMDet-S Link to heading

Those two times are different machines and different lengths. In the recipes, 12 epochs is called 1× and 36 epochs is called 3×. The fair pair is in the v0.4.0 note: on one RTX 3090 Ti, RTMDet-S 3× finishes in 8h 58m (15.3 tiles/s) and RetinaNet OBB 3× in 23h 23m (5.9 tiles/s).

Finetune RTMDet from this 3× checkpoint. It is the published RTMDet weight, and the same file holds both the AP50 peak and the AP75 peak. Oriented R-CNN at 3× reaches an AP75 of 51.23.

The three ResNet models Link to heading

Use this table when you are finetuning a ResNet backbone, or when the scene is the bus lot and you want a head we have already looked at on that picture.

Finetune these three from the 1× checkpoint. On them, a 3× train leaves AP50 flat or lower. What moves is how tight the box is. The full table is the v0.2.0 zoo. RTMDet’s numbers are the v0.4.0 note.

Oriented R-CNN’s lead here is the tight box. On AP50 the three sit within about 2–4 points. On AP75 the gap opens to about 8–10 points (50.24 vs 41.90 vs 40.40).

Finding the object is a separate question, and the three are close. On validation tiles that were also in the training set, recall over every stored detection is 93.1% / 91.8% / 92.8%. Recall at the score where hits and false alarms balance is 83.5% / 83.0% / 83.6%. During training, the share of real objects covered at half overlap runs the other way: Faster R-CNN 88.6%, FCOS 87.2%, Oriented R-CNN 76.1%. These figures use half overlap. A looser overlap of one tenth is absent from the published tables.

When the job is to find the object, and you are staying on ResNet, pick Faster R-CNN or FCOS. Pay for Oriented R-CNN when the box has to be right and the scene looks like the bus lot. On AP75 the published peak is RTMDet-S at 52.01.

Oriented R-CNN, when the row is diagonal Link to heading

This is the pick inside the ResNet three when the box has to follow the vehicle.

On large vehicles, the class that looks like the bus lot, its DOTA AP50 is 82.54, against Faster R-CNN’s 75.25. On the held-out ship test the gap shrinks (90.41 vs 88.77 vs 88.34) and the boxes sit closer to the hull: mean best overlap 0.85, against 0.79 and 0.75.

The second step is why. Oriented R-CNN proposes regions that are already rotated, then cuts a crop turned to the object (oriented RoIAlign). The rest of the network sees the vehicle lined up in that crop. On the diagonal bus lot that is one box per vehicle, along the row. Faster R-CNN on the same tile draws a pile. The bus-lot check found the same pile in official MMRotate, so the cause is this design.

That second crop is also why the train is heavy. The published 1× recipe takes 1 day 12 hours on one L4: mixed precision off, batch 2, a full score every 4 epochs, 13,691 tiles. As shipped, that recipe does not fit my RTX 3090. The Faster R-CNN finetunes in this series (SSDD, HRSID) ran on an RTX 3090 Ti.

Three changes before a bigger card Link to heading

The file on the Hub is the accuracy recipe. These three decide whether Oriented R-CNN starts on a 24 GB card. They apply to the other ResNet models too. RTMDet’s published recipe already uses batch 8 on a 3090 Ti.

Turn mixed precision on. training.use_amp is false on the published recipes. odet train --use-amp overrides it. The training guide notes about 2× the step rate and a lower memory bill. In practice the accuracy has held. The day and a half above is the time with mixed precision off.

Raise the batch. data_loader.batch_size is 2, and --batch-size overrides it. That batch leaves memory free. Raise it while the step still fits. The learning rate in the file was chosen for batch 2, so a larger batch is a new run, with a shorter clock.

Score less during training. A validation pass still runs every epoch. The slow part is matching every polygon to compute the score, on evaluation.val_map_every_n_epochs (4 on DOTA 1×). Set that to 0 to skip the match, or set dataset.max_val_samples to score fewer tiles. When the match itself is the cost, raise evaluation.val_score_threshold (the recipes use 0.3). The trainer suggests this once a pass keeps more than 50,000 boxes. That floor changes the in-training best_mAP. odet preds and the published number keep a floor of 0.05.

Faster R-CNN for most ResNet finetunes Link to heading

This is the ResNet model I actually finetune. DOTA AP50 74.42%, ships 88.77%, about 11 h 30 m for DOTA 1× on the same L4, and about 1.15× the tiled inference speed of Oriented R-CNN (7.565 vs 6.604 images/s on the 7,669-tile stitch, one RTX 3090 Ti). FAIR1M, SSDD, and HRSID in v0.3 all start from rotated_faster_rcnn_dota_le90_1x.

Aircraft fits this model. Plane AP50 on DOTA is 89.40% (FCOS 88.74). On the optical demo, Faster R-CNN 3× keeps the full Pleiades apron (27 aircraft) and the dense Tucson storage field (125 planes). Headings follow the fuselage. A plane is long, and it can sit near 45°, with space around it. A bus row is packed.

The miss is specific: long objects, packed together, near 45°. Faster R-CNN proposes an upright region and cuts an upright crop, then predicts a rotated box from that crop. On an isolated ship or aircraft the crop is mostly the object. On a chevron of buses the crop includes the neighbours, and the angle comes out wrong. Same tile, score at least 0.6, NMS 0.1 (boxes that overlap by more than a tenth are merged):

Ships in HRSC are long. Faster R-CNN is 1.6 points behind Oriented R-CNN there. I use it for aircraft, for ordinary ship scenes, and for scenes where the objects have space around them.

FCOS for that diagonal row, on a shorter train Link to heading

RTMDet is a one-step model too, and it leads DOTA. FCOS is the one-step ResNet head we have checked on the bus lot, with a published time on the L4.

The row. On the bus lot, FCOS boxes follow the chevron. It draws each box in one step from the feature map. The macOS walkthrough uses the same demo.jpg. Official DOTA AP50 is 73.07%. On large vehicles the gap to Faster R-CNN is small (76.05 vs 75.25). What you buy is the heading on that row, on a lighter train than Oriented R-CNN.

The clock. One head, no proposal step, no rotated crop. On the L4 the 1× train takes 7 h 59 m, against 11 h 30 m for Faster R-CNN and a day and a half for Oriented R-CNN. The 3× ship train has the same shape: 34 m, 1 h 5 m, 2 h 34 m. RTMDet’s 8h 58m is 36 epochs on a 3090 Ti.

FCOS writes its score with a sigmoid, so useful scores sit lower. Start the cutoff at 0.20. A cutoff of 0.60, copied from Faster R-CNN, hides the scene.

Links Link to heading

• oriented-det on GitHub · docs · pretrained zoo
• Numbers: v0.4.0 RTMDet · Task 1 zoo · HRSC · ProbIoU / Faster R-CNN
• The exception: bus lot, MMRotate and OrientedDet · FCOS vs Oriented R-CNN · optical demo, aircraft
• Read the number: le90 · Oriented mAP
• Previous: Does MMRotate Faster R-CNN also mess up the bus lot?

Written on October 29, 2026 by Jeff Faudi. Link to heading