System Learning Deep Learning (39)--ohem



Training region-based Object detectors with Online hard Example Mining INTRO:CVPR. Online Hard Example Mining (Ohem) arxiv:http://arxiv.org/abs/1604.03540 paper:http://www.cv-foundation.org/ Openaccess/content_cvpr_2016/papers/shrivastava_training_region-based_object_cvpr_2016_paper.pdf GitHub (official): Https://github.com/abhi2610/ohem author page:http://abhinav-shrivastava.info/

Track and transfer:watching Videos to simulate strong Human supervision for weakly-supervised Object Detection I NTRO:CVPR arxiv:http://arxiv.org/abs/1604.05766

Exploit all the layers:fast and accurate CNN Object Detector with scale Dependent Pooling and cascaded rejection Clas Sifiers intro:scale-dependent Pooling (SDP), cascaded rejection Clas-sifiers (CRC) paper:http:// Www-personal.umich.edu/~wgchoi/sdp-crc_camready.pdf

Transferred from: http://blog.csdn.net/u012905422/article/details/52760669

This paper presents an efficient target detection algorithm that trains region-based convolution detection operators by online hard example Mining (Ohem) algorithm, which can suppress simple samples and some small numbers of samples, and make the training process more efficient. This method makes use of the significant bootstrapping technique (commonly used in SVM) to modify the SGD algorithm, so that the original region-based convnets heuristic learning and multi-parameters can be removed, and the results are more accurate and stable. The maps in Pascal VOC2007 and 2012 are: 78.9%, 76.3%, respectively.
Hard Example Mining:

There are 2 main methods of the hard example mining algorithm, optimizing the SVM time algorithm and non-SVM utilization.

When using hard example mining in optimized SVM, the training algorithm mainly maintains the balanced iterative process of training SVM and convergence on the working set, while removing some work-set samples and adding other special standards during the update process. The standard here is to remove some easily distinguishable sample classes and add some sample classes that cannot be judged by the existing model for new training. The working set is a small piece of data throughout the training set.

When used in non-SVM, the hard example mining algorithm starts with a positive sample data set and a random negative sample dataset, which is trained in these datasets to converge on the dataset and apply it to other untrained negative sample sets. The negative sample data (false positives) is added to the training set and the model is trained again. This process is usually iterative only once, and does not get a lot of training convergence process.
Network Structure framework:

The Ohem algorithm is improved based on fast R-CNN algorithm, and the author thinks that the fast R-CNN algorithm is not efficient and optimal when it is used to create mini-batch for SGD algorithm, and Ohem can obtain lower training loss, and higher MAP. Compare the two algorithms of fast R-CNN and Ohem structure:

It can be found that the Ohem algorithm presented in the paper, for a given image, after selective search RoIs, also calculates the convolution feature graph. In the green section (a), however, a read-only ROI network propagates forward the feature map and all ROI, and the Hard ROI module uses the loss of these ROI to select B samples. In the red section (b), these selected samples (hard examples) Enter the ROI network for further forward and back propagation.
Experimental Results:

The experiment results are very good, only the map results are attached here: