[1] Xiangning Chen, Cihang Xie, Mingxing Tan, Li Zhang, Cho-Jui Hsieh, Boqing Gong. CVPR 2021. Arxiv link: https://arxiv.org/abs/2103.13886
Det-AdvProp is a data augmentation technique specifically designed for the fine-tuning process of object detectors. It can consistently and substantially outperform the vanilla training and AutoAugment under various settings. The obtained detector is not only more accurate on clean images, but also more robust to image distortions and domain shift.
The following table includes a list of models trained with Det-AdvProp + AutoAugment (AA):
| Model | APtest | AP50 | AP75 | APS | APM | APL | APval | #params | #FLOPs | |
|---|---|---|---|---|---|---|---|---|---|---|
| EfficientDet-D0 + Det-AdvProp + AA (ckpt, test-dev) | 35.3 | 54.1 | 37.8 | 12.7 | 39.9 | 53.2 | 35.1 | 3.9M | 2.54B | |
| EfficientDet-D1 + Det-AdvProp + AA (ckpt, test-dev) | 40.9 | 60.0 | 44.1 | 19.1 | 45.6 | 57.2 | 40.8 | 6.6M | 6.10B | |
| EfficientDet-D2 + Det-AdvProp + AA (ckpt, test-dev) | 44.3 | 63.5 | 47.9 | 23.5 | 48.5 | 59.9 | 44.3 | 8.1M | 11.0B | |
| EfficientDet-D3 + Det-AdvProp + AA (ckpt, test-dev) | 48.0 | 67.1 | 52.2 | 28.1 | 51.8 | 62.8 | 47.7 | 12.0M | 24.9B | |
| EfficientDet-D4 + Det-AdvProp + AA (ckpt, test-dev) | 50.4 | 69.5 | 54.9 | 30.9 | 54.3 | 64.4 | 50.4 | 20.7M | 55.2B | |
| EfficientDet-D5 + Det-AdvProp + AA (ckpt, test-dev) | 52.5 | 71.8 | 57.2 | 34.6 | 55.9 | 65.2 | 52.2 | 33.7M | 130B |
Unlike the vanilla EfficientDet that scales the image with mean and std, here we scale the input to the range of [-1, 1] to make it easier for performing adversarial attack. Please see this Colab for reproducing the results.
We test the detectors' robustness against common image corruptions (e.g., Gaussian Noise, Snow, etc.) based on the COCO-C dataset in this paper. The table below shows the comparison between vanilla training and Det-AdvProp + AutoAugment (AA):
| Model | mAP |
|---|---|
| EfficientDet-D0 | 21.4 |
| + Det-AdvProp + AA | 22.7 (+1.3) |
| EfficientDet-D1 | 24.4 |
| + Det-AdvProp + AA | 26.7 (+2.3) |
| EfficientDet-D2 | 26.7 |
| + Det-AdvProp + AA | 28.9 (+2.2) |
| EfficientDet-D3 | 28.8 |
| + Det-AdvProp + AA | 32.0 (+3.2) |
| EfficientDet-D4 | 30.1 |
| + Det-AdvProp + AA | 33.9 (+3.8) |
| EfficientDet-D5 | 31.4 |
| + Det-AdvProp + AA | 35.0 (+3.6) |
PASCAL VOC 2012 only contains 20 classes, which are much smaller than the 80 labeled classes in COCO. The underlying distributions of the two datasets are also different in the image content or the bounding box sizes and locations. We use the trained detectors to run inference directly on the VOC dataset to test their transferibility. We maintain the COCO evaluation metrics in this experiment:
| Model | mAP | AP50 | AP75 |
|---|---|---|---|
| EfficientDet-D0 | 55.6 | 77.6 | 61.4 |
| + Det-AdvProp + AA | 56.2 (+0.6) | 78.3 (+0.7) | 62.3 (+0.9) |
| EfficientDet-D1 | 60.8 | 82.0 | 66.7 |
| + Det-AdvProp + AA | 61.3 (+0.5) | 82.5 (+0.5) | 67.6 (+0.9) |
| EfficientDet-D2 | 63.3 | 83.6 | 69.3 |
| + Det-AdvProp + AA | 63.6 (+0.3) | 84.0 (+0.4) | 70.0 (+0.7) |
| EfficientDet-D3 | 65.7 | 85.3 | 71.8 |
| + Det-AdvProp + AA | 66.4 (+0.7) | 85.9 (+0.6) | 72.8 (+1.0) |
| EfficientDet-D4 | 67.0 | 86.0 | 73.0 |
| + Det-AdvProp + AA | 67.8 (+0.8) | 87.0 (+1.0) | 74.3 (+1.3) |
| EfficientDet-D5 | 67.4 | 86.9 | 73.8 |
| + Det-AdvProp + AA | 68.7 (+1.3) | 88.0 (+1.1) | 75.4 (+1.6) |