Configs

Training and evaluation are driven by JSON config files with MMRotate-style _base_ inheritance. Base paths may be file-relative (same directory as the recipe), absolute, or prefixed with @odet: (path relative to the oriented-det repository root).

DOTA pretrained models (model zoo)

OrientedDet publishes DOTA le90 pretrain checkpoints on Hugging Face Hub. See pretrained/README.md and per-model READMEs below.

Split protocol (important):

• Training: train+val — train_tiles_dirs unions train and val tile roots (DOTA benchmark pretrain, same idea as MMRotate).
• Evaluation: val — mAP is on the val tile set only.
• This is not a fine-tune train/val holdout: val tiles are included in training.

mAP in README tables uses make eval-val mAP50 (all 7,669 val tiles, filter_empty_gt=false, rotated IoU ≥ 0.50). Training-time periodic mAP may be higher (non-empty tiles only).

Model	Config	Hub slug	eval-val mAP50
Oriented R-CNN 1×	oriented_rcnn/dota_le90_1x.json	oriented_rcnn_dota_le90_1x	74.79%
Oriented R-CNN 3×	oriented_rcnn/dota_le90_3x.json	oriented_rcnn_dota_le90_3x	79.40%
Rotated RetinaNet 1×	rotated_retinanet/dota_le90_1x.json	rotated_retinanet_dota_le90_1x	64.14%
Rotated RetinaNet 3×	rotated_retinanet/dota_le90_3x.json	rotated_retinanet_dota_le90_3x	71.52%
Rotated Faster R-CNN 1×	rotated_faster_rcnn/dota_le90_1x.json	rotated_faster_rcnn_dota_le90_1x	77.57%
Rotated Faster R-CNN 3×	rotated_faster_rcnn/dota_le90_3x.json	rotated_faster_rcnn_dota_le90_3x	83.42%

Download: odet pretrained download <slug> or "load_from_checkpoint": "hf://<slug>". Published eval-val reports: docs/eval-reports/ (markdown + analysis JSON; predictions.json stays in gitignored predictions/ for the viewer).

Prefix	Resolves under	Override env
@odet:	ORIENTED_DET_ROOT (repo root)	ORIENTED_DET_ROOT, or installed package location

Example in an external project that keeps local dataset fragments:

{
  "_base_": [
    "my_dataset.json",
    "@odet:configs/_base_/models/oriented_rcnn_r50.json",
    "@odet:configs/_base_/schedules/1x.json"
  ]
}

Documentation

• Configuration reference — human-readable guide: sections, recipes, CLI, production vs evaluation
• config.schema.json — machine-readable types and defaults (kept in sync with oriented_det/train/config.py)

Layout

• base/ — Dataset, model, schedule, FP16, preprocessing, augmentation fragments (do not run directly; included via _base_).
• oriented_rcnn/ — Oriented R-CNN (horizontal RPN + midpoint-offset + oriented ROI). See oriented_rcnn/README.md.
• rotated_faster_rcnn/ — Rotated Faster R-CNN. Standard recipe: ProbIoU main ROI loss, Smooth L1 aux 0.1, angle weight 1.0 (dota_le90_1x.json; 3× via dota_le90_3x.json). See rotated_faster_rcnn/README.md.
• rotated_retinanet/ — Rotated RetinaNet (one-stage). See rotated_retinanet/README.md.

Muted keys

Prefix a key with _muted_ to keep an alternate value in the file without affecting training (stripped before validation):

{
  "training": {
    "learning_rate": 0.002,
    "_muted_learning_rate": 0.005
  }
}

Full option reference

All options, types, and defaults: config.schema.json (synced with oriented_det/train/config.py and tools/train.py). Human-readable detail: Configuration reference.

Key order (schema, saved config.json, docs): switches first (lr_scheduler_type, loss_type, enable_albumentation), then grouped prefixes (lr_*, roi_*, rpn_*). Top-level sections:

Section	Description
_base_	Base config path(s): relative to current file, @odet:…, or absolute
model_type	oriented_rcnn, rotated_faster_rcnn, or rotated_retinanet
dataset	data_root, format (dota / airbus_playground), train_tiles_dir, val_tiles_dir, train_tiles_dirs, val_tiles_dirs (optional lists; union without on-disk merge), same_folder (DOTA: images and .txt in same dir), overlap (tile overlap px, even; deploy uses margin = overlap/2), annotations_file, split_file, val_split_id, train_includes_val (Airbus: train on all folds; val fold for monitoring only), difficult_strategy, filter_empty_gt (DOTA: drop tiles with no GT after filters; MMRotate parity), max_train_samples, max_val_samples, max_samples_shuffle_seed (deterministic spread when capping), allowed_classes, ignore_labels, map_labels
data_loader	batch_size, num_workers, shuffle, pin_memory
model	backbone, fpn_, anchor_, target_means/stds, roi_* (loss, batch, iou, schedule, roi_proj_xy), rpn_, use_hbb_for_matching, add_gt_as_proposals, rpn_nms_threshold (proposal NMS), final_nms_iou_threshold / final_nms_iou_schedule_ (post–ROI-head NMS), final_nms_use_cpu (exact polygon final NMS on CPU), nms_class_agnostic, max_detections_per_image, inference_pre_nms_score_threshold
training	lr_scheduler_type first, then lr_scheduler_, lr_warmup_steps, lr_scaling_, use_lr_param_groups, lr_mult_*, then num_epochs, learning_rate, momentum, weight_decay, use_amp, gradient_accumulation_steps, max_grad_norm, loss_weights. See Training — Learning rate scheduling and base/schedules/README.md.
evaluation	score_threshold, iou_threshold, compute_map_final, compute_map_every_n_epochs
production	Optional overrides: val/mAP — score_threshold overrides evaluation when set; per_class_score_threshold merges on top of evaluation (mAP IoU is evaluation.iou_threshold only). Decode (checkpoint inference only) — RPN/NMS/threshold fields patch the loaded model via apply_inference_config_to_model in load_model_from_checkpoint (deploy, save_predictions, image_demo); not applied during tools/train.py (training uses model.* only). Deploy / tiling — overlap_pixels, ignore_margin_pixels, canvas flags (see config.schema.json).
checkpoint	load_from_checkpoint, load_from_experiment, discover_previous_run, resume_from_checkpoint_epoch, load_optimizer_state, load_scheduler_state, load_include_prefixes, load_exclude_prefixes, start_epoch, best_metric, higher_is_better
loss	loss_type, class_weight_method, background_weight, focal_alpha, focal_gamma, label_smoothing, roi_grouped_ce_* (coarse-to-fine ROI classifier curriculum in one run)
preprocessing	resize_mode, target_size, normalize_mean, normalize_std, pad_size_divisor, enable_flip_horizontal, enable_flip_vertical, enable_flip_diagonal
Top-level	enable_albumentation, enable_profiling
augmentation	Albumentations params (when enable_albumentation is true)
tensorboard	log_debug_anchors_proposals, vis_score_threshold

PyPI package (oriented_det.configs)

A subset of this tree is copied into oriented_det/configs/ for wheels (see oriented_det/configs/vendored_manifest.txt), including Rotated RetinaNet and Rotated Faster R-CNN DOTA recipes. After editing files here:

make sync-configs    # update vendored copy
make check-configs   # verify before commit (also runs in CI)

Training

From the repo root:

python tools/train.py --config configs/oriented_rcnn/dota_le90_1x.json

Override batch size or AMP: --batch-size, --use-amp, --no-amp. See tools/README.md and the main README.