Add New Data Augmentation Technique: OA-Mix

[dazory]

Motivation

This pull request introduces OA-Mix, a key component of the OA-DG (Object-Aware Domain Generalization) framework, designed to improve domain generalization in object detection. The goal is to enhance the model's robustness against domain shifts by generating diverse multi-domain data without compromising object annotations.

OA-Mix was first introduced in the paper titled "Object-Aware Domain Generalization for Object Detection," which was presented as an oral session at AAAI 2024. The method has shown state-of-the-art performance on domain generalization benchmarks for object detection.

Modification

This PR implements the following:

• Implemented the OA-Mix class with configurable parameters.
• Integration of OA-Mix into the existing data pipeline for seamless use during model training.
• Added test functions to validate the effectiveness and stability of the OA-Mix method.
• Updated the level_to_mag() function to prevent the level from being divided into only two steps when min_mag=0.0 and max_mag=0.1 in translation transformations by changing round(*, 1) to round(*, 2).

BC-breaking (Optional)

The proposed modification is backward-compatible. OA-Mix is introduced as an optional augmentation technique, and the update to level_to_mag() only improves the precision of existing transformations, without breaking previous usage.

If you want to use OAMix, simply modify the train_pipeline in the configuration file as shown below:

train_pipeline = [
    dict(type='LoadImageFromFile', backend_args=None),
    dict(type='LoadAnnotations', with_bbox=True),
    dict(
        type='RandomResize',
        scale=[(2048, 800), (2048, 1024)],
        keep_ratio=True),
    dict(type='RandomFlip', prob=0.5),
    dict(type='OAMix')
    dict(type='PackDetInputs')
]

Use cases (Optional)

OA-Mix is particularly useful for:

• Scenarios where models need to generalize to unseen domains, such as varying weather conditions or image corruptions in real-world environments.
• Object detection tasks requiring high fidelity in localization and classification under domain shifts, especially when only a single-source domain is available for training.
• Documentation and example usage have been updated to demonstrate how OA-Mix can be integrated into different object detection models.

Below is a performance comparison between a baseline object detection model and the same model with OA-Mix applied:

Model	Dataset	Backbone	mAP	mPC	Gauss.	Shot	Impulse	Defocus	Glass	Motion	Zoom	Snow	Frost	Fog	Bright	Contrast	Elastic	Pixel	JPEG
Faster R-CNN	Cityscapes-C	R-50-FPN	42.2	15.4	0.5	1.1	1.1	17.2	16.5	18.3	2.1	2.2	12.3	29.8	32.0	24.1	40.1	18.7	15.1
Faster R-CNN + OA-Mix	Cityscapes-C	R-50-FPN	42.7	20.8	7.2	9.6	7.7	22.8	18.8	21.9	5.4	5.2	23.6	37.3	38.7	31.9	40.2	22.2	20.2

The model was evaluated using the robust detection benchmark, which can be run using the test_robustness.py script provided by mmdetection.

Checklist

• Pre-commit tools have been run, and all issues have been resolved.
• Comprehensive unit tests have been added to ensure the correctness and reproducibility of the OA-Mix implementation.
• The modification has been tested with MMDet, confirming compatibility and improved performance.
• Documentation has been updated accordingly, including docstrings and example tutorials demonstrating how to use OA-Mix.

[CLAassistant]

All committers have signed the CLA.

[dazory]

Hi @jbwang1997, I just wanted to kindly remind you to review this pull request when you get a chance. Your feedback would be greatly appreciated. 😊 Thanks in advance!