Improving object detection with boosted histograms

TLDR: This work shows how histogram-based image descriptors can be combined with a boosting classifier to provide a state of the art object detector and introduces a weak learner for multi-valued histogram features and shows how to overcome problems of limited training sets.

About: This article is published in Image and Vision Computing.The article was published on 2009-04-01 and is currently open access. It has received 127 citations till now. The article focuses on the topics: Object-class detection & Viola–Jones object detection framework.

Figures

Table 1 Average precision for object detection on test1 VOC 2005 image set

Table 2 Average precision for object detection on test2 VOC 2005 image set

Fig. 1. Rectangles on the left and right image are examples of possible regions for histogram features. Stable appearance in A, B and C on both images makes corresponding features to be good candidates for a motorbike classifier. On the contrary, regions D are unlikely to contribute for the classification due to the large variation in appearance.

Fig. 8. Examples of true positive (TP) and false positive (FP) detections of motorbikes and (Top) FP motorbike detections (red) frequently correspond to bicycles. (Bottom) Many hig classified as FP due to the insufficient overlap with (partly missing) annotation rectangl

Fig. 3. Histogram features. (Left) Sample regions from an exhaustive set of regions defined by different spatial extents and positions within the object window. (Right) Histogram features computed for each region according to four types of spatial grids.

Fig. 2. Classification of two diagonal distributions using (left): AdaBoost with weak learners in terms of axis-parallel linear classifiers; (right): Fisher linear discriminant.

Frequently asked questions

Q1. What are the contributions in "Improving object detection with boosted histograms" ?

Building upon recent progress in the field the authors show how histogram-based image descriptors can be combined with a boosting classifier to provide a state of the art object detector. Among the improvements the authors introduce a weak learner for multi-valued histogram features and show how to overcome problems of limited training sets. The authors also analyze different choices of image features and address computational aspects of the method. In particular, using a single set of parameters their approach outperforms all the methods reported in VOC05 Challenge for seven out of eight detection tasks and four object classes while providing close to real-time performance. 

Q2. What is the method for training the classifier?

To train the classifier for a particular object class, the authors use positive training set with scale and positionnormalized images of objects in similar views. 

Q3. What is the advantage of using WFLD as an AdaBoost weak learner?

Using WFLD as an AdaBoost weak learner eliminates the need of re-sampling training data required by classifiers that do not make use of sample weights. 

Q4. What is the advantage of using a weak learner?

A particular advantage of using FLD as a weak learner is the possibility of re-formulating FLD to minimize a weighted classification error as required by AdaBoost. 

Q5. How can The authorfind an efficient classifier for n training samples?

For one-dimensional features f 2 R such as Haar features in [29], an efficient classifier for n training samples can be found by selecting an optimal decision threshold in (1) in O(n logn) time. 

Q6. What is the cost of a limited reduction of performance?

Given the high correlation of filter responses at adjacent image scales, computation of integral histograms for a limited set of sparse scale levels is likely to imply a speed up at the cost of a limited decrease of performance. 

Q7. What is the way to deal with multi-dimensional features?

One approach to deal with multi-dimensional features used in [15] is to project f onto a pre-defined set of one-dimensional manifolds using a fixed set of functions gj : Rm ! 

Q8. What is the way to test the proposed method?

To validate the proposed method, the authors test it on the task of object detection in natural images and evaluate the performance on PASCAL Visual Object Category datasets VOC 2005 and VOC 2006 [7,6]. 

Q9. What is the procedure used to evaluate the detection of motorbikes?

To train and to test the detectors the authors use training and validation sets of VOC 2005 challenge and adopt VOC evaluation procedure [7]. 

Q10. How does the method perform on different object classes?

the authors find Harris-Affine regions to perform no better than random regions in their framework tested on three different object classes.