Training Region-based Object Detectors with Online Hard Example Mining

最新推荐文章于 2022-04-07 13:32:38 发布

原创最新推荐文章于 2022-04-07 13:32:38 发布 · 449 阅读

0 ·

CC 4.0 BY-SA版权

文章标签：

#OHEM #hard #example #mining

deep-learning 专栏收录该内容

20 篇文章

订阅专栏

本文介绍了一种名为OHEM（Online Hard Example Mining）的方法，该方法通过自动选择困难样本简化了训练过程并提高了效率。OHEM消除了传统训练中常用的几个启发式方法和超参数，特别是在基于区域的卷积神经网络（ConvNets）中。实验结果显示，随着训练集规模的增大及难度提升，OHEM能够显著提高平均精度。

摘要生成于 C知道，由 DeepSeek-R1 满血版支持，前往体验 >

DOI: CVPR 2016

Key Qustion

the training set is distinguished by a large imbalance between the number of annotated objects and the number of background examples

Contribution

Make training more effective and efficient.
OHEM is a simple and intuitive algorithm that eliminates several heuristics and hyperparameters in common use.
The candidate examples are subsampled according to a distribution that favors diverse, high loss instances.
It removes the need for several heuristics and hyperparameters commonly used in region-based ConvNets.
It yields a consistent and significant boosts in mean average precision
Its effectiveness increases as the training set becomes larger and more difficult, as demonstrated by results on the MS COCO dataset

Architecture

Fast R-CNN

这里写图片描述

OHEM

这里写图片描述

Experiments

这里写图片描述

Conclusion

OHEM eliminates several heuristics and hyperparameters in common use by automatically selecting hard examples, thus simplifying training.
Though we used Fast R-CNN throughout this paper, OHEM can be used for training any region-based ConvNet detector.

Unknown Key Words

bootstrapping = hard negative mining rely on aforementioned alternation template:(a) for some period of time a fixed model is used to find new examples to add to the active training set; (b) then, for some period of time the model is trained on the fixed active training set.
hard positive example = false positive example

Questions

However, there is a small caveat: co-located RoIs with high overlap are likely to have correlated losses.
- use standard non-maximum suppression (NMS) to perform deduplication
*

Self-Learning

SGD is not suitable for bootstrapping template
2 methods of hard example mining
- remove easy example and then add some hard example
- add false positives to dataset to train the model again
proposal’s IOU with ground truth is in the interval [bg_lo, 0.5), bg_lo = 0.1 is helpful but ignore some infrequent, but import, difficult background regions.
OHEM is robust in case one needs fewer images per batch in order to reduce GPU memory usage.