Learning Data Augmentation Strategies for Object Detection

CoRR, 2019.

Cited by: 105|Views348
We find that a learned data augmentation policy is effective across all data sizes considered, with a larger improvement when the training set is small


Data augmentation is a critical component of training deep learning models. Although data augmentation has been shown to significantly improve image classification, its potential has not been thoroughly investigated for object detection. Given the additional cost for annotating images for object detection, data augmentation may be of ev...More



Full Text
  • Deep neural networks are powerful machine learning systems that work best when trained on vast amounts of data.
  • Recent work has shown that instead of manually designing data augmentation strategies, learning an optimal policy from data can lead to significant improvements in generalization performance of image classification models [22, 45, 8, 33, 31, 54, 2, 43, 37, 5].
  • We investigate how the performance of a data augmentation policy depends on the number of operations included in the search space and how the effective of the augmentation technique varies as dataset size changes.
  • Design and implement a search method to combine and optimize data augmentation policies for object detection problems by combining novel operations specific to bounding box annotations.
  • A good data augmentation policy is one that can transfer between models, between datasets and work well for models trained on different image sizes.
  • We experiment with the learned augmentation policy on a different backbone architecture and detection model.
  • These experiments show that the augmentation policy transfers well across a different backbone architecture, detection algorithm, image sizes (i.e. 640 → 1280 pixels), and training procedure .
  • To evaluate the transferability of the learned policies to an entirely different dataset and another different detection algorithm, we train a Faster R-CNN [39] model with a ResNet-101 backbone on PASCAL VOC dataset [11].
  • The improvements due to the learned augmentation policy is larger when the model is trained on smaller datasets, which can be seen in Fig. 3 and in Table 5.
  • As the training set size is increased, the effect of the learned augmentation policy is decreased, the improvements are still significant.
  • It is interesting to note that models trained with learned augmentation policy seem to do especially well on detecting smaller objects, especially when fewer images are present in the training dataset.
  • For small objects, applying the learned augmentation policy seems to be better than increasing the dataset size by 50%, as seen in Table.
  • 5. For small objects, training with the learned augmentation policy with 9000 examples results in better performance than the baseline when using 15000 images.
  • When we apply the learned data augmentation, the training loss is increased significantly for all dataset sizes.
  • We investigate the application of a learned data augmentation policy on object detection performance.
  • We find that a learned data augmentation policy is effective across all data sizes considered, with a larger improvement when the training set is small.
  • We show that for models trained on 5,000 training samples, the learned augmentation policy can improve mAP by more than 70% relative to the baseline
Your rating :