FineDOTA — DOTA v1.0 / v1.5 in the unified detection format

Source: the Ultralytics-hosted DOTA release archives (github.com/ultralytics/assets, DOTAv1 / DOTAv1.5 / DOTAv2), which bundle the original DOTA-format annotations (labels/*_original) used for this conversion. Images are the original resolution, re-encoded by upstream from PNG to JPEG.

Converted by the finedet project into a unified, AutoTrain-compatible layout: image / width / height / objects{bbox, category} with COCO-format [x, y, w, h] boxes in absolute pixels. Boxes are clipped to the image and empty boxes dropped; category ids are densified per the category tables below.

This repository hosts one config per source release: v1.0, v1.5 (default: v1.0). Load one with load_dataset("finedet/dota", "<config>").

Box format

objects.bbox follows the COCO convention: [x, y, w, h] in absolute pixels, origin at the image's top-left corner.

Schema notes

Each object carries three box representations. Which is original and which is derived:

• objects.polygon — the original annotation: the raw DOTA 8-point oriented box [x1, y1, x2, y2, x3, y3, x4, y4] in absolute pixels, vertex order exactly as in the source label file (DOTA annotates the four vertices in clockwise order, starting from a category-specific starting vertex), not clipped to the image bounds.
• objects.bbox — derived: the axis-aligned hull of the polygon ([min x, min y, max x - min x, max y - min y] over the 4 points), then clipped to the image; an object whose clipped hull has zero width or height is dropped entirely (polygon and obb included).
• objects.obb — derived: [cx, cy, w, h, theta], the minimum-area rectangle over the 4 polygon points (cv2.minAreaRect), normalized to the long-edge convention: w >= h, theta = angle of the w edge against the +x axis, in radians within [-pi/2, pi/2) (for squares the two equivalent angles are not disambiguated). Not clipped.
• objects.difficult — the original DOTA difficult flag (kept, not filtered).

The sample galleries below draw the derived axis-aligned bbox hulls.

License

Original dataset: DOTA (all releases share the same terms). Official terms: 'All images and their associated annotations in DOTA can be used for academic purposes only, but any commercial use is prohibited.' The same terms apply to the redistributed archives used as the source here. Imagery originates from Google Earth, CycloMedia B.V., and the GF-2 / JL-1 satellites (panchromatic), and remains subject to the respective providers' terms — the Google Earth terms of use in particular. Academic use only.

Config v1.0

Example images

Boxes are colored by category: near-transparent fill, opaque outline.

Conversion notes

Converted from the Ultralytics DOTA v1.0 archive using the bundled original-format annotations. 15 categories; train 1,411 / validation 458 images. difficult boxes are kept and flagged via objects.difficult. No tiling — original image sizes are preserved (up to ~20,000 px per side); image files are upstream JPEG re-encodes of the official PNGs. The test / test-dev / test-challenge splits are excluded (annotations not public).

Splits

• train: 1411 images
• validation: 458 images

Categories

id	name
0	plane
1	ship
2	storage-tank
3	baseball-diamond
4	tennis-court
5	basketball-court
6	ground-track-field
7	harbor
8	bridge
9	large-vehicle
10	small-vehicle
11	helicopter
12	roundabout
13	soccer-ball-field
14	swimming-pool

Config v1.5

Example images

Boxes are colored by category: near-transparent fill, opaque outline.

Conversion notes

Converted from the Ultralytics DOTA v1.5 archive using the bundled original-format annotations. 16 categories (adds container-crane); the same images as v1.0, relabeled — including very small instances. difficult boxes are kept and flagged via objects.difficult. No tiling — original image sizes are preserved (up to ~20,000 px per side); image files are upstream JPEG re-encodes of the official PNGs. The test / test-dev / test-challenge splits are excluded (annotations not public).

Splits

• train: 1411 images
• validation: 458 images

Categories

id	name
0	plane
1	ship
2	storage-tank
3	baseball-diamond
4	tennis-court
5	basketball-court
6	ground-track-field
7	harbor
8	bridge
9	large-vehicle
10	small-vehicle
11	helicopter
12	roundabout
13	soccer-ball-field
14	swimming-pool
15	container-crane

Training with transformers

The boxes are already in the absolute-pixel COCO [x, y, w, h] format that AutoImageProcessor expects, so fine-tuning a detector needs no bbox conversion:

import torch
from datasets import load_dataset
from transformers import (AutoImageProcessor, AutoModelForObjectDetection,
                          Trainer, TrainingArguments)

ds = load_dataset("finedet/dota", "v1.0")
obj_feat = ds["train"].features["objects"]
if hasattr(obj_feat, "feature"):
    obj_feat = obj_feat.feature
cat_feat = obj_feat["category"]
names = (cat_feat.feature if hasattr(cat_feat, "feature") else cat_feat).names

checkpoint = "facebook/detr-resnet-50"
processor = AutoImageProcessor.from_pretrained(checkpoint)
model = AutoModelForObjectDetection.from_pretrained(
    checkpoint,
    id2label=dict(enumerate(names)),
    label2id={n: i for i, n in enumerate(names)},
    ignore_mismatched_sizes=True,
)


def transform(batch):
    images = [img.convert("RGB") for img in batch["image"]]
    annotations = [
        {"image_id": i,
         "annotations": [
             {"bbox": box, "category_id": cat, "area": box[2] * box[3], "iscrowd": 0}
             for box, cat in zip(objs["bbox"], objs["category"])
         ]}
        for i, objs in enumerate(batch["objects"])
    ]
    return processor(images=images, annotations=annotations, return_tensors="pt")


def collate(batch):
    return {"pixel_values": torch.stack([x["pixel_values"] for x in batch]),
            "labels": [x["labels"] for x in batch]}


trainer = Trainer(
    model=model,
    args=TrainingArguments(output_dir="out", per_device_train_batch_size=4,
                           num_train_epochs=10, learning_rate=1e-5,
                           remove_unused_columns=False),
    train_dataset=ds["train"].with_transform(transform),
    data_collator=collate,
)
trainer.train()