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.
The checkpoints were trained on DOTA (academic-use). The seven scenes are a mix of DOTA tiles and commercial optical rasters used as research illustration — not a commercial product. For production detectors on your own imagery, contact DL4EO.
What is on the page Link to heading
Three Hub 3× slugs, ResNet-50 + FPN, DOTA le90. Same protocol as the published zoo: train+val pretrain, eval on all 7,669 val tiles.
| Model | Architecture | eval-val mAP50 | Demo threshold | Hub slug |
|---|---|---|---|---|
| Rotated Faster R-CNN | two-stage, horizontal RPN | 83.46% | 0.70 | rotated_faster_rcnn_dota_le90_3x |
| Rotated FCOS | one-stage, anchor-free | 82.32% | 0.25 | rotated_fcos_dota_le90_3x |
| Oriented R-CNN | two-stage, oriented RPN | 79.40% | 0.75 | oriented_rcnn_dota_le90_3x |
The default is Rotated Faster R-CNN. Switching models also resets the slider to that checkpoint’s DOTA eval-val best-F1 threshold. That is the same trick as the macOS FCOS walkthrough: do not copy 0.70 onto FCOS, or you will drop half the boxes.
Seven scenes:
| Scene | Sensor / source | What it stresses |
|---|---|---|
| DOTA vehicles 1024 | DOTA tile | dense diagonal buses and trucks |
| DOTA vehicles large | DOTA tile | same class mix, larger canvas |
| Pleiades Airport | Pleiades | aircraft at mixed headings |
| Pleiades HD15 Miami Marina | Pleiades | packed ships, harbors, a few vehicles |
| Pleiades Neo Tucson | Pleiades Neo | dense aircraft storage |
| SPOT Storage | SPOT | circular tanks |
| NZ marina | optical | long piers, many headings |
Vehicles: the two-stage models agree, FCOS agrees on geometry Link to heading
The 1024 DOTA bus lot is the same family of scene as demo.jpg from the June macOS post. At each model’s operating point the counts sit on top of each other: 98 (Faster R-CNN) / 97 (Oriented R-CNN) / 101 (FCOS). Headings follow the chevron parking. The visible difference is score calibration, not box shape — FCOS spreads confidence across roughly 0.4–0.9; the two-stage heads pile many boxes near 1.00.
On the large DOTA vehicle tile the same pattern holds (112 / 109 / 115). If you only look at those two scenes, you could believe the three detectors are interchangeable. The other five images are there to stop that.
Planes, ships, tanks Link to heading
Airport. All three Oriented-Det models keep 27 aircraft on the Pleiades apron — identical count, boxes that follow fuselage heading. That is the easy scene, and it is in the demo so you can see a clean overlay before the marinas get busy.
Tucson. A denser aircraft field. Faster R-CNN and FCOS both keep 115 planes (plus a handful of helicopters); Oriented R-CNN is 113. The overlay is the one I would send someone who still thinks axis-aligned boxes are “good enough” for aircraft:
Miami marina. After a tight merge NMS (IoU 0.1), ship counts line up: 258 / 252 / 269. Harbor is the class that still splits the one-stage head from the two-stage heads (13 vs 6–7). FCOS is a little hungrier on piers; the R-CNN models are leaner.
NZ marina. Same story at smaller GSD. FCOS keeps the most ships (322 plus 7 harbors); Oriented R-CNN the fewest (256 ships). Faster R-CNN sits in the middle (271).
SPOT storage. Compact circular tanks are a published FCOS strength on eval-val. On this tile the three Oriented-Det models are close (87 / 83 / 89 tanks at their operating points), which is what you want from a demo: no one is quietly failing a class the zoo said they could do.
How the three Oriented-Det models compare to each other Link to heading
I matched boxes 1-to-1 at oriented IoU ≥ 0.5 with the same class, after each model’s best-F1 threshold, on all seven images together.
| Pair | Matched | F1 | Median IoU |
|---|---|---|---|
| Faster R-CNN vs Oriented R-CNN | 1,024 | 0.98 | 1.00 |
| FCOS vs Faster R-CNN | 983 | 0.89 | 0.87 |
| FCOS vs Oriented R-CNN | 960 | 0.89 | 0.87 |
Oriented R-CNN at 0.75 is essentially a strict subset of Faster R-CNN at 0.70: every Oriented R-CNN box matches, median IoU 1.00, median angle difference 0°. The 36 unmatched Faster R-CNN boxes are the 0.70–0.75 score band. After the NMS 0.1 recompute, the two two-stage heads are no longer drawing different Miami extras — they are drawing the same objects.
FCOS is the one that still looks like a different detector. Pairwise F1 against either R-CNN is 0.89, median IoU 0.87. That is strong agreement on where the objects are, with the residual coming from elongated classes (ships, harbors) and from FCOS keeping a few more boxes in dense marinas.
Totals at operating points, all seven images:
| Class | Faster R-CNN | Oriented R-CNN | FCOS |
|---|---|---|---|
| ship | 529 | 508 | 591 |
| large-vehicle | 162 | 160 | 166 |
| plane | 142 | 140 | 142 |
| small-vehicle | 111 | 105 | 108 |
| storage-tank | 87 | 83 | 89 |
| harbor | 7 | 6 | 20 |
| kept (all classes) | 1,060 | 1,024 | 1,139 |
In parity with MMRotate Link to heading
MMRotate is the research reference — see the v0.1.1 parity notes and the ProbIoU post. I did run a reference MMRotate Rotated Faster R-CNN on the same seven images, same matching rule, at its own best-F1 threshold (0.70).
The short version: the three Oriented-Det checkpoints are comparable to that MMRotate run. They do not look like a different product category.
| Pair | Matched | F1 | Median IoU |
|---|---|---|---|
| MMRotate vs FCOS | 965 | 0.89 | 0.87 |
| MMRotate vs Faster R-CNN | 900 | 0.86 | 0.86 |
| MMRotate vs Oriented R-CNN | 880 | 0.86 | 0.86 |
Kept-box counts sit in the same band: 1,021 (MMRotate) vs 1,060 / 1,024 / 1,139. Ship totals are 559 vs 529 / 508 / 591. Vehicles and tanks are within a few boxes of each other. On these scenes, Oriented-Det’s three 3× checkpoints land in parity with a reference MMRotate Rotated Faster R-CNN, with the residual concentrated on a few dense, out-of-DOTA rasters rather than on the DOTA-like tiles.
Latency, for orientation only Link to heading
These numbers are inference duration stored in the prediction JSON — useful as a rough ordering, not as a benchmark. Mean over the seven scenes: FCOS 1.42 s, Faster R-CNN 2.47 s, Oriented R-CNN 2.42 s. The two R-CNN heads now sit in the same latency band (Miami is ~7 s for both). The MMRotate reference run is in the same order of magnitude (mean 1.13 s on this machine and this tiling).
For a proper Apple Silicon timing table, the FCOS vs Oriented R-CNN macOS note is the better source. The demo page does not re-run the models.
How to read the slider Link to heading
| Detector | Start here | Why |
|---|---|---|
| Rotated Faster R-CNN | 0.70 | peaked two-stage scores; the page default |
| Oriented R-CNN | 0.75 | even more peaked; below ~0.70 you mostly add duplicates the NMS already removed |
| FCOS | 0.25 | sigmoid head; 0.70 will hide most of the scene |
If you want a fair visual comparison, leave each model on its default threshold, then switch the radio buttons. Dragging all three to the same number is how you convince yourself FCOS is “worse.”
Reproduce the overlays Link to heading
The page is static, but the boxes come from the same CLI as every other post in this series:
odet pretrained download rotated_faster_rcnn_dota_le90_3x
odet pretrained download rotated_fcos_dota_le90_3x
odet pretrained download oriented_rcnn_dota_le90_3x
odet image-demo path/to/scene.jpg hf://rotated_faster_rcnn_dota_le90_3x \
--score-thr 0.70 --nms-thr 0.1 --out-file out_rfrcnn.png
Swap the slug and threshold for FCOS (0.25) or Oriented R-CNN (0.75). Keep --nms-thr 0.1 unless you have a reason to match a looser two-stage production config.
Links Link to heading
- Demo: dl4eo.com/object-detection-optical-satellite
- oriented-det on GitHub · PyPI · docs
- Pretrained zoo
- Previous: Rotated FCOS vs Oriented R-CNN on macOS · Oriented-Det v0.2.0 · v0.1.1 / MMRotate parity