OpticsBench, LensCorruptions
收藏资源简介：
OpticsBench和LensCorruptions数据集用于研究深度神经网络在图像分类和目标检测任务中的鲁棒性。OpticsBench包含主要光学像差作为图像失真，如彗差、散光和球差，而LensCorruptions则包含从100个真实镜头中获得的混合像差。这些数据集通过追踪光线根据惠更斯原理，并结合泽尼克多项式进行空间分辨和参数化，以模拟真实的光学像差。评估了超过70种不同的视觉模型在应用了这些失真的ImageNet数据集上的表现，结果表明当前模型的鲁棒性不足。此外，提出了一种名为OpticsAugment的数据增强方法，该方法在OpticsBench上平均提高了18%的性能，并允许将学到的鲁棒性转移到其他领域。
The OpticsBench and LensCorruptions datasets are designed to investigate the robustness of deep neural networks (DNNs) across image classification and object detection tasks. OpticsBench incorporates core optical aberrations as image distortions, including coma, astigmatism, and spherical aberration, whereas LensCorruptions features mixed aberrations collected from 100 real-world lenses. These two datasets simulate realistic optical aberrations by tracing light rays in accordance with Huygens' principle and leveraging Zernike polynomials for spatial resolution and parameterization. Over 70 distinct vision models were evaluated on the ImageNet dataset corrupted by these distortions, and the experimental results demonstrate that current state-of-the-art visual models exhibit insufficient robustness. Furthermore, a data augmentation method named OpticsAugment is proposed, which achieves an average 18% performance improvement on OpticsBench and enables the transfer of learned robustness to other downstream domains.
数据集概述：Classification robustness to common optical aberrations
数据集基本信息
- 作者: Patrick Müller
- 年份: 2023
- 许可证: 参见/opticsbench和/opticsaugment目录
- 相关论文: Classification Robustness to Common Optical Aberrations
数据集内容
- OpticsBench: 用于研究对现实光学模糊效果的鲁棒性基准，包含由Zernike多项式导出的光学像差（如coma、astigmatism、spherical、trefoil）。
- OpticsAugment: 一种使用光学核的数据增强方法，可提高模型对光学像差的鲁棒性。
主要功能
-
生成预定义的图像损坏数据集: python python benchmark.py --generate_datasets --database imagenet-1k_val --testdata_path <path_to_validation_images>
-
评估PyTorch DNNs: python python benchmark.py --run_all --path_to_root_folder <root> --models all
-
使用OpticsAugment训练模型: python python train_dnn.py --root_dir <path_to_dataset> --model_dir $path_to_modeldir --name <model_name> --num_workers <num_workers>
数据集结构
root/ images/ <dataset>/ /val /corruptions <corruption_name>/ <severity>/ eval/ <dataset>/ /val /corruptions <corruption_name>/ <severity>/ <model_name>.json models/ <model_checkpoints>.pt
性能指标
-
OpticsBench ImageNet-100上的准确率（平均所有损坏）:
- DenseNet (ours): 68.22 | 65.33 | 56.33 | 41.60 | 30.13
- EfficientNet (ours): 61.00 | 55.34 | 42.14 | 30.27 | 23.35
- MobileNet (ours): 57.59 | 52.30 | 38.58 | 27.51 | 20.54
- ResNet101 (ours): 69.90 | 67.68 | 61.36 | 49.04 | 37.80
- ResNeXt50 (ours): 65.14 | 62.68 | 54.44 | 39.90 | 28.45
-
在2D常见损坏上的性能提升（平均差异，%-points）:
- DenseNet161: 5.08 | 7.55 | 8.73 | 7.30 | 5.38
- ResNeXt50: 5.11 | 7.63 | 8.68 | 7.18 | 5.27
- ResNet101: 1.25 | 3.07 | 4.55 | 4.90 | 4.10
- MobileNet: 3.58 | 4.92 | 4.78 | 3.69 | 3.07
- EfficientNet: 4.35 | 6.32 | 6.70 | 4.62 | 3.69
引用
bibtex @InProceedings{Muller_2023_ICCV, author = {M"uller, Patrick and Braun, Alexander and Keuper, Margret}, title = {Classification Robustness to Common Optical Aberrations}, booktitle = {Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV) Workshops}, month = {October}, year = {2023}, pages = {3632-3643} }
