Object Detection with discriminatively trained Part Based Models

P. felzenszwalb, R. girshick, D. mcallester, D. ramanan
Object Detection with discriminatively trained Part Based Models
IEEE Transactions on Pattern Analysis and machine intelligence, vol. 32, No. 9, SEP. 2010

Read this article, not because of DPM, but because of training the hard negative mining of SVM. This article introduces hard negative mining for SVM and hard negative mining for latent SVM respectively. How to Select a negative sample when training SVM classifier? The author suggests that the negative samples should be negative samples with incorrect scores, that is, the samples on the margin edge of SVM should not be, and the negative samples with correct scores should be easily avoided, what we need is those that have the ability to distinguish, but are also divided into errors. In the SVM training process, samples are divided during each training. Easy negative sample E and hard negative sample H. Before each training process starts, first, clean the training sample and remove the E sample from the negative sample. If necessary, add the h sample. The samples on the margin and the samples on the margin are not counted as E or H. The whole is to continuously optimize SVM by selecting the samples that are difficult to distinguish, the difference between common SVM and lsvm is that the sample selection method is the same for different loss functions. The sample selection process is as follows,

Let C1 d be an initial cache of examples.
Algorithm repeatedly trains a model and updates
Cache as follows:
1) Let T: = (CT) (train a model using CT ).
2) if h (T; d) CT stop and return T.
3) Let c0t: = ctnx for any X such that x e (t; CT)
(Shrink the cache ).
4) Let CT + 1: = c0t [X for any X such that x D and X \ h (T; d) NCT 6 =; (grow the cache ).

The negative sample selection of SVM is also the same idea in R-CNN. PS. Listen to Shuicheng Yan's report. They also perform SVM training in the same way.

 

To be continued ......

 

Object Detection with discriminatively trained Part Based Models