Add oriented bounding box support for DOTAv1 datasets

[janerikvw]

Summary

This PR adds comprehensive support for oriented bounding boxes (OBB) in RF-DETR, including a new DOTAv1 dataset loader, multi-dataset configuration system, and OBB-specific operations and loss functions.

Key Changes

New Dataset Support

• DOTAv1 Dataset Loader (src/rfdetr/datasets/dota.py): Implements DotaDetection class to load DOTAv1 format datasets with 8-point corner annotations. Supports both oriented and axis-aligned bounding box modes.
• Multi-Dataset Configuration (src/rfdetr/datasets/multi.py): Adds YAML-based multi-dataset configuration system with ClassMappingDataset wrapper for class remapping and OBB padding across heterogeneous datasets.

OBB Operations

• Box Operations (src/rfdetr/utilities/box_ops.py):
  • corners_to_obb(): Converts 4 corner points to normalized OBB representation (cx, cy, w, h, angle)
  • obb_to_corners(): Inverse conversion from OBB to corner points
  • circular_angle_loss(): Circular loss for angle predictions accounting for 180° symmetry

Model Updates

• LWDETR (src/rfdetr/models/lwdetr.py): Added oriented parameter to predict 5-dim boxes (cx, cy, w, h, angle) when enabled
• Criterion (src/rfdetr/models/criterion.py): Added oriented flag and loss_angle_coef for angle loss weighting
• PostProcess (src/rfdetr/models/postprocess.py): Updated to handle 5-dim oriented box outputs
• Matcher (src/rfdetr/models/matcher.py): Updated to handle oriented boxes in cost computation

Configuration & Training

• Config (src/rfdetr/config.py):
  • Added oriented field to ModelConfig
  • Added loss_angle_coef to TrainConfig
  • Extended dataset_file to support "dota" and "multi" formats
  • New DatasetEntry and MultiDatasetConfig classes for multi-dataset YAML configs
• Transforms (src/rfdetr/datasets/transforms.py): Enhanced Normalize to convert OBB corners to normalized angle representation; updated geometric transforms to handle corner points as keypoints
• Data Module (src/rfdetr/training/module_data.py): Added weighted sampling support for multi-dataset training
• Namespace (src/rfdetr/_namespace.py): Propagated oriented flag through build pipeline

Testing

• DOTAv1 Tests (tests/datasets/test_dota.py): Comprehensive tests for dataset loading, oriented/non-oriented modes, class mapping, and edge cases
• Multi-Dataset Tests (tests/datasets/test_multi.py): Tests for YAML config parsing, class mapping, and OBB padding
• OBB Operations Tests (tests/utilities/test_obb_box_ops.py): Tests for corner↔OBB conversions, roundtrip consistency, and circular angle loss
• Config Tests (tests/test_obb_config.py): Tests for new configuration fields

Implementation Details

• OBB angles are normalized to [0, π) to handle 180° symmetry of rectangles
• Corner points are treated as keypoints during augmentation to preserve orientation
• Multi-dataset support includes per-dataset class mapping and augmentation config overrides
• Weighted sampling enables balanced training across datasets with different sizes

https://claude.ai/code/session_01DC76DdnusAVRPbF1NXfska

[CLAassistant]

Thank you for your submission! We really appreciate it. Like many open source projects, we ask that you sign our Contributor License Agreement before we can accept your contribution.
_{You have signed the CLA already but the status is still pending? Let us recheck it.}

[codecov]

Codecov Report

❌ Patch coverage is 59.68170% with 152 lines in your changes missing coverage. Please review.
✅ Project coverage is 74%. Comparing base (3cd6898) to head (8670cd6).
⚠️ Report is 4 commits behind head on develop.

❌ Your patch check has failed because the patch coverage (60%) is below the target coverage (95%). You can increase the patch coverage or adjust the target coverage.
❌ Your project check has failed because the head coverage (74%) is below the target coverage (95%). You can increase the head coverage or adjust the target coverage.

Additional details and impacted files

@@           Coverage Diff           @@
##           develop   #842    +/-   ##
=======================================
- Coverage       75%    74%    -1%     
=======================================
  Files           92     94     +2     
  Lines         7148   7509   +361     
=======================================
+ Hits          5383   5594   +211     
- Misses        1765   1915   +150     

🚀 New features to boost your workflow:

• ❄️ Test Analytics: Detect flaky tests, report on failures, and find test suite problems.

[Borda]

Thank you for your submission! We really appreciate it. Like many open source projects, we ask that you sign our Contributor License Agreement before we can accept your contribution.

@janerikvw ^^ 🙏