ACF目标检测算法论文.pdf

发布时间：2022-05-31 发布人：admin 分类：说明书资料大小：2.47M 资料格式：pdf 举报版权申诉

gxl10281-9941399-16359647507495560803.pdf-第1页.png

第1页 / 共8页

gxl10281-9941399-16359647507495560803.pdf-第2页.png

第2页 / 共8页

gxl10281-9941399-16359647507495560803.pdf-第3页.png

第3页 / 共8页

gxl10281-9941399-16359647507495560803.pdf-第4页.png

第4页 / 共8页

gxl10281-9941399-16359647507495560803.pdf-第5页.png

第5页 / 共8页

gxl10281-9941399-16359647507495560803.pdf-第6页.png

第6页 / 共8页

gxl10281-9941399-16359647507495560803.pdf-第7页.png

第7页 / 共8页

gxl10281-9941399-16359647507495560803.pdf-第8页.png

第8页 / 共8页

1 . Introduction

2 . Related work

3 . Proposed face detector

3.1 . Feature description

3.2 . Investigation guidelines

3.3 . Feature design

3.4 . Training design

4 . Multi-view detection

4.1 . View partition

4.2 . Post-processing

5 . Experiments

5.1 . Evaluation on benchmark face database

5.2 . Discussion

6 . Conclusion

4 1 0 2 p e S 3 ] V C . s c [ 2 v 3 2 0 4 . 7 0 4 1 : v i X r a Aggregate Channel Features for Multi-view Face Detection Bin Yang Junjie Yan Zhen Lei Stan Z. Li∗ Center for Biometrics and Security Research & National Laboratory of Pattern Recognition Institute of Automation, Chinese Academy of Sciences, China {jjyan,zlei,szli}@nlpr.ia.ac.cn yb.derek@gmail.com Abstract Face detection has drawn much attention in recent decades since the seminal work by Viola and Jones. While many subsequences have improved the work with more pow- erful learning algorithms, the feature representation used for face detection still can’t meet the demand for effectively and efﬁciently handling faces with large appearance vari- ance in the wild. To solve this bottleneck, we borrow the concept of channel features to the face detection domain, which extends the image channel to diverse types like gradi- ent magnitude and oriented gradient histograms and there- fore encodes rich information in a simple form. We adopt a novel variant called aggregate channel features, make a full exploration of feature design, and discover a multi- scale version of features with better performance. To deal with poses of faces in the wild, we propose a multi-view detection approach featuring score re-ranking and detec- tion adjustment. Following the learning pipelines in Viola- Jones framework, the multi-view face detector using ag- gregate channel features shows competitive performance against state-of-the-art algorithms on AFW and FDDB test- sets, while runs at 42 FPS on VGA images. 1. Introduction Human face detection have long been one of the most fundamental problems in computer vision and human- computer interaction. In the past decade, the most inﬂuen- tial work should be the face detection framework proposed by Viola and Jones [22]. The Viola-Jones (abbreviated as VJ below) framework uses rectangular Haar-like features and learns the hypothesis using Adaboost algorithm. Com- bined with the attentional cascade structure, the VJ detector achieved real-time face detection at that time. Despite the great success of the VJ detector, the performance is still far from satisfactory due to the large appearance variance of faces in unconstrained settings. ∗Corresponding author. Figure 1. An intuitive visualization of our multi-view face detec- tor using aggregate channel features. The area with warmer color indicates more attention paid to by the detector. To handle faces in the wild, many subsequences of VJ framework merged. These methods mainly get the performance gains in two aspects, more complicated fea- tures [17, 19, 26] and (or) more powerful learning algo- rithms [14, 1, 25]. As the combination of boosting and cas- cade has been proven to be quite effective in face detection, the bottleneck lies in the feature representation since com- plicated features adopted in the above literatures bring about limited performance gains at the cost of large computation cost. Lately in another domain of pedestrian detection, a fam- ily of channel features has achieved record performances [6, 5]. Channel features compute registered maps of the orig- inal images like gradients and histograms of oriented gra- dients and then extract features on these extended chan- nels. The classiﬁer learning process follows the VJ frame- work pipeline. In this paper, we adopt a variant of chan- nel features called aggregate channel features [5], which are extracted directly as pixel values on subsampled chan- nels. Channel extension offers rich representation capac- ity, while simple feature form guarantees fast computation. With these two superiorities, the aggregate channel features break through the bottleneck in VJ framework and have the potential to make great advance in face detection.

As we mainly concentrate our efforts to the feature rep- resentation rather than learning algorithms in this paper, we not only just adopt the aggregate channel features in face de- tection, but also try to explore the full potential of this novel representation. To do so, we make a deep and all-round investigation into the speciﬁc feature parameters concern- ing channel types, feature pool size, subsampling method, feature scale and so on, which gives insights into the fea- ture design and hopefully provides helpful guidelines for practitioners. Through the deep exploration, we ﬁnd that: 1) multi-scaling the feature representation further enriches the representation capacity since original aggregate channel features have uniform feature scale; 2) different combina- tions of channel types impact the performance greatly, while for face detection the color channel in LUV space, plus gra- dient magnitude channel and gradient histograms channels in RGB space show best result; 3) multi-view detection is proven to be a good match with aggregate channel features as the representation naturally encodes the facial structure (Figure 1). Although multi-view detection could effectively deal with diverse poses, additional issues come up as how to merge detections output by separately trained subview de- tectors, and how to deal with the offsets of location and scale between output detections and ground-truth. We solve these problems by carefully designed post-processing in- cluding score re-ranking, detection merging and bounding box adjustment. The detailed experimental exploration of aggregate channel features, along with our improvements on multi- view detection, leads to large performance gain in face de- tection in the wild. On two challenging face databases, AFW and FDDB, the proposed multi-view face detector shows competitive performance against state-of-the-art de- tectors in both detection accuracy and speed. The remaining parts of this paper are organized as fol- lows. Section 2 revisits related work in face detection. Sec- tion 3 describes how we build the face detector using aggre- gate channel features. Section 4 addresses problems con- cerning multi-view face detection. Experimental results on AFW and FDDB are shown in section 5 and we conclude the paper in section 6. 2. Related work Face detection has drawn much attention since the early time of computer vision. Although many solutions had been put forward, it was not until Viola and Jones [22] pro- posed their milestone work that face detection saw surpris- ing progress in the past decades. The VJ face detector fea- tures in three aspects: fast feature computation via integral image representation, classiﬁer learning using Adaboost, and the attentional cascade structure. One main drawback of the VJ framework is that the features have limited repre- sentation capacity, while the feature pool size is quite large to compensate for that. Typically, in a 24 × 24 detection window, the number of Haar-like features is 160,000 [22]. To address the problem, efforts are made in two directions. Some focus on more complicated features like HoG [26], SURF [13]. Some aim to speed up the feature selection in a heuristic way [18, 2]. However, the problem hasn’t been solved perfectly. In this paper, we mainly focus on the fea- ture representation part and make a deep exploration into it, which is complementary to existing work on the learning algorithm and classiﬁer structure in the VJ framework. Recently channel features have been proposed and shown record performance in pedestrian detection [6, 5]. Due to the channel extension to diverse types like gradi- ents and local histograms, the features show richer repre- sentation capacity for classiﬁcation. However, the features are extracted as rectangular sums at various locations and scales which we believe leads to a redundant feature pool. During preparation of this paper, Mathias et al. [16] inde- pendently discover the effectiveness of integral channel fea- tures in face detection domain. In this paper, we adopt a novel variant of channel features called aggregate channel features, which extract features directly as pixel values in extended channels without computing rectangular sums at various locations and scales. The feature has powerful rep- resentation capacity and the feature pool size is only several thousands. Through careful design in section 3 and imple- mentation of multi-view detection in section 4, the aggre- gate channel features based detector achieves state-of-the- art performance on challenging databases. 3. Proposed face detector In this section, we make a full exploration of the aggre- gate channel features in the context of face detection. We ﬁrst give a brief introduction of the feature itself, including its computation, properties and advantages over traditional Haar-like features used in VJ framework. Then the detailed experimental investigation is described in two parts, feature design and training design. Before that, some guidelines concerning how we conduct the investigation are demon- strated. Each design part is divided into several separate ex- periments ended with a summary explaining the speciﬁc pa- rameters used in our proposed face detector. Note that each experiment focuses on only one parameter and the others remain constant. Through the well-designed experiments, the proposed face detector based on aggregate channel fea- tures is built step by step. Issues concerning the implemen- tation of multi-view face detection which further improves the performance are discussed in the next section. 3.1. Feature description Channel extension: The basic structure of the aggre- gate channel features is channel. The application of channel

lar sums with various locations and scales using integral images, leading to a faster feature computation and smaller feature pool size for boosting learning. With the help of cas- cade structure, detection speed is accelerated more. 3) Due to its structure consistence with the overall image, when coupled with boosting method, the boosted classiﬁer nat- urally encodes structured pattern information from large training data (see Figure 1 for an illustration), which gives more accurate localization of faces in the image. 3.2. Investigation guidelines All investigations are trained on the AFLW face database1 [10] and tested on the Annotated Faces in the Wild (AFW) testset2. To make it clear, there are in total 36, 112 positive samples and 108, 336 negative samples se- lected from AFLW which are kept constant in all investiga- tions. Testset contains 205 natural images with faces that vary a lot in pose, appearance and illumination. To alleviate the ground-truth offset caused by different annotation styles (Figure 4) in training and testing set and make the evaluation more comparable, a lower Jaccard in- dex3 with threshold 0.3 is adopted in comparative evalu- ation. Practically the lower threshold won’t cause errors being mistakenly corrected. Note that in ﬁnal evaluation of the proposed face detector (section 5), the AFW test- set, together with another face benchmark FDDB database, are used as testbed and the evaluation metric follows the database protocol. 3.3. Feature design To fully exploit the power of aggregate channel features in face detection domain, a deep investigation into the de- sign of the feature is done mainly on channel types, win- dow size, subsampling method and feature scale. Results of comparative experiments are shown in Figure 6. Channel types: Three types of channels are used, which are color channel (Gray-scale, RGB, HSV and LUV), gra- dient magnitude, and gradient histograms. The computation of the latter two channel types could be seen as a general- ized version of HoG features. Speciﬁcally, gradient magni- tude is the biggest response on all three color channels, and oriented gradient histograms follow the idea of HoG in that: 1) rectangular cell size in HoG equals the subsampling fac- tor in aggregated channel features; 2) each orientation bin results in one feature channel (6 orientation bins are used in this paper). Figure 6 (a)˜(c) show how much each of these three types alone contributes to the performance of face detection. It can be seen that the gradient histograms contribute most to the performance among all three channel 1http://lrs.icg.tugraz.at/research/aflw/ 2http://www.ics.uci.edu/˜xzhu/face/ 3The Jaccard index is deﬁned as the size of the intersection divided by the size of the union of the sample sets. Figure 2. Work-ﬂow of proposed face detector. has a long history since digital images were invented. The most common type of channel should be the color chan- nels of the image, with Gray-scale and RGB being typical ones. Besides color channels, many different channel types have been invented to encode different types of informa- tion for more difﬁcult problems. Generally, channels can be deﬁned as a registered map of the original image, whose pixels are computed from corresponding patches of original pixels [6]. Different channels can be computed with linear or non-linear transformation of the original image. To al- low for sliding window detection, the transformations are constrained to be translationally invariant. Feature computation: Based on the deﬁnition of chan- nels, the computation of aggregate channel features is quite simple. As shown in Figure 2, given a color image, all deﬁned channels are computed and subsampled by a pre- set factor. The aggregate pixels in all subsampled channels are then vectorized into a pixel look-up table. Note that an optional smoothing procedure can be done on each chan- nel with a binomial ﬁlter both before computation and after subsampling. Classiﬁer learning: The learning process is quite sim- ple. Two changes are made compared with VJ framework. First is that weak classiﬁer is changed from decision stump to depth-2 decision tree. The more complex weak classiﬁer shows stronger ability in seeking the discriminant intra and inter channel correlations for classiﬁcation [15]. Second difference is that soft-cascade [1] structure is used. Unlike the attentional cascade structure in VJ framework which has several cascade stages, a single-stage classiﬁer is trained on the whole training data and a threshold is then set after each weak classiﬁer picked by Adaboost. These two changes lead to more efﬁcient training and detection. Overall superiority: Compared with traditional Haar- like features used in VJ framework, aggregate channel fea- tures have the following differences and advantages: 1) The image channels are extended to more types in order to en- code diverse information like color, gradients, local his- tograms and so on, therefore possess richer representation capacity. 2) Features are extracted directly as pixel values on downsampled channels rather than computing rectangu-

types. Figure 6 (d) shows the performances of combinations of these three types computed on different color channels. Detection window size: Detection window size is the scale to which we resize all face and non-face samples and then train our detector. Larger window size includes more pixels in feature pool and thus may improve the face detec- tion performance. On the other hand, too large window will miss some small faces and diminish the detection efﬁciency. Figure 6 (e) shows comparison of window size ranging from 32 to 112 with a stride of 16 pixels. Subsampling: The factor for subsampling can be re- garded as the perceptive scale for that it controls the scale at which the aggregation is done. Changing the factor from large to small leads to the feature representation shifting from coarse to ﬁne and the feature pool size getting bigger. Experiments on different subsampling factors are shown in Figure 6 (f). In original aggregate channel features, the way to do subsampling is average pooling. Following the idea in Convolutional Neural Networks, another two ways of subsampling, max pooling and stochastic pooling [24] are tested in Figure 6 (g). Smoothing: As described in feature description, both pre and post smoothing is done in default setting of aggre- gate channel features. A binomial ﬁlter with a radius of 1 is used for smoothing. The smoothing procedure also has a great inﬂuence on the scale of the feature represen- tation. Concretely, pre-smoothing determines how far the local neighborhood is in which local correlations are en- coded before channel computation, while post-smoothing determines the neighborhood size in which the computed channel features are integrated with each other. In [6], the former corresponds to the ‘local scale’ of the feature, while the latter represents the ‘integration scale’. We vary the ﬁl- ter radius used in pre and post smoothing and ﬁnd that both using a radius of 1 gets the best results. Figure 6 (h)˜(i) present the comparative results. Multi-scale: In aggregate channel features, although hidden information at different scale could be extracted at a cost of more weak classiﬁers, it would be better to make the integrated channel features multi-scaled and thus make themselves more discriminant. Therefore the same or bet- ter classiﬁcation performance can be achieved with fewer weak classiﬁers. In this part, we implement three multi- scale version of aggregate channel features in the afore- mentioned three kinds of scale, perceptive scale (subsam- pling), local scale (pre-smoothing) and integration scale (post-smoothing) and compare their performaces. See re- sults in Figure 6 (j)˜(l). Summary: The color channel, gradient magnitude and gradient histograms prove themselves a good match in ag- gregate channel features. However, different choices of color channel used and on which gradients are computed have a great impact on performance. According to the ex- periments, LUV channel and gradient magnitude and 6-bin histograms computed on RGB color space (in total 10 chan- nels) are the best choice for face detection. Larger detection window size generally gets better per- formance, but will miss many small faces in testing and lead to inefﬁcient detection. In this work, we set the size to 80 × 80 as its optimal performance. A subsampling factor of 4 is most reasonable according to the experiments, while different pooling methods show small differences. However, max pooling and stochastic pooling are much slower than average pooling, therefore the average pooling becomes the best match for the sake of ef- ﬁciency. In this way, the resulting feature pool size of our face detector is (80/4)× (80/4)× 10 = 4000, considerably smaller than that in VJ framework. As for multi-scale version of aggregate channel fea- tures, multi-local-scale with an additional scale of radius 2 shows the best performance. The probable reason is that pre-smoothing controls the local scale of the neighborhood feature correlations and therefore matches the intuition in- side multi-scale best. Compared with other ﬁne-tuning, the multi-scale version has a notable performance gain for that it makes up for the scale uniformity caused by subsampling to some extent. One main drawback is that it doubles the feature pool size and as a result slows down the detection speed somewhat. Based on the trade-off, we implement two face detectors with different scale settings, one is single- scaled with faster speed and the other is multi-scaled with better accuracy. We evaluate and discuss the performances of these two versions in detail in section 5. 3.4. Training design Besides careful design of the aggregate channel features, experiments on the training process which is similar to that in VJ framework are also carried out. The differences are that the weak classiﬁer is changed into depth-2 decision tree and soft-cascade [1] structure is used. Details of the training design are as follows. Number of weak classiﬁers: Given a feature pool size of 4, 000, we vary the number of weak classiﬁers con- tained in the soft-cascade. In Figure 3 performances of various numbers of weak classiﬁers ranging from 32 to 8192 are displayed, which shows that apparently more clas- siﬁers generate better performance, and when the number gets larger the performance begins to saturate. Since more classiﬁers slow down the detection speed, there’s a trade- off between accuracy and speed. Searching for the saturate point as the optimal is signiﬁcant during training in such framework. Training data: Empirically, more training data will get better performance given powerful representation capacity. In this case, AFLW database is used as the only positive training data. However, as images in AFLW database are

Figure 3. Comparison of different numbers of weak classiﬁer in the soft cascade. very salient and the background has very less variance, neg- ative samples cropped from the AFLW database can’t rep- resent the real world scenario well, which limits the face detection performance in the wild. In this part, we further use PASCAL VOC 2007 database and randomly crop win- dows from images without person as the new negative sam- ples. Experiments show that the new training data contain- ing cluttered background signiﬁcantly improve the perfor- mance with 4.1%. Summary: Based on observations above, we choose 2048 as the number of weak classiﬁers contained in the soft cascade. As each weak classiﬁer is a depth-2 decision tree, it takes only two comparing operations to apply a weak clas- siﬁer, which is quite fast. During training, as negative data is large, we adopt a standard Bootstrap procedure to sam- ple hard negative samples from PASCAL VOC 2007 in the implementation of the proposed face detector. 4. Multi-view detection Human faces in real world usually have highly varied poses. In AFLW database, the human pose is divided into three aspects: 1 in-plane rotation ‘roll’ and 2 out-of-plane rotations ‘yaw’ and ‘pitch’. Because of this large variance in face pose, it is difﬁcult to train a single view face detec- tor to handle all the poses effectively. A multi-view detec- tion is further examined in this part. Due to the adoption of soft-cascade structure, a multi-view version of face detec- tor won’t cause too much computation burden. Typically, we divide the out-of-plane rotation yaw into different views and let the classiﬁer itself tolerate the pose variance in the other two types of rotations. Adopting multi-view detection also brings about many troublesome issues. If handled improperly, the performance will differ greatly. First, detectors of different view will each produce a set of candidate positive windows followed with a set of conﬁdence scores. For application purpose, we need to merge these detections from different views and also remove duplicated windows. A typical approach is Non- Maximum Suppression (NMS) [3]. An issue rises on how to compare conﬁdence scores from different classiﬁers and how to do window merging in the trade-off between high Figure 4. Illustration of different ground-truth annotation styles in databases, the view partition and symmetric detection adjustment. Rectangles with red and green color correspond to detections be- fore and after adjustment. precision rate and high detection rate. Second, as for de- tection evaluation, usually the overlap of bounding boxes is used as the criterion. However, annotations in different data sets may not have a consistent style (Figure 4 (a)). This di- versity suffers more in proﬁle faces. Since our face detector is trained and tested on different data sets, this issue impacts the performance a lot. Third, detectors of different views need to be trained with different samples separately. How to divide the views therefore becomes another concerning problem. In this section, we address the above three issues successfully by careful designs and therefore fully exploit the advantage of multi-view detection. 4.1. View partition In the scenario of detecting faces in the wild, pose vari- ation caused by yaw is usually severer than pitch and roll. Therefore we divide the faces in AFLW database according to yaw angle. We have 6 subviews which are horizontally symmetric (see Figure. 4 (b)) because we ﬂip each image in the training set. Speciﬁcally, there are 6630, 8446, 9610, 9610, 8446, 6630 images in views from 1 to 6. Beneﬁtting from the symmetry of our model, we can only train three subview detectors of the right side for simplicity, and use these trained right-side detectors to generate the left-side detectors. Detections of all six detectors are then merged to get the ﬁnal detections. Though multi-view detection sig- niﬁcantly improves the detection performance (especially the recall rate), the post-processing of detections from dif- ferent detectors becomes a trouble. If handled improperly, the performance degrades a lot. 4.2. Post-processing Difﬁculties in the post-processing of multi-view detec- tion mainly reﬂect on the following aspects: 1) different score distributions and; 2) different bounding box styles.

Concretely, as each subview detector is trained separately, their output conﬁdence scores usually have different dis- tributions. What’s more, due to the annotation rule in the AFLW database that the face’s nose is approximately at the center location of the bounding box ground-truth, as the subview changes, the bounding box shifts. This bound- ing box offset causes difﬁculty both in detection merging and ﬁnal evaluation using Jaccard index metric. To solve these annoying issues and make the best use of multi-view detection, we introduce the following methods for post- processing. Score re-ranking: We propose the following three kinds of score re-ranking: 1) normalizing scores of different views to [0, 1]; 2) deﬁning a new score that has uniform distribution and; 3) taking overlapping detections into con- sideration. N ormalization: After training a classiﬁer, calculate the output range of the classiﬁer and use the range to do nor- malization later so that output score has a range of [0, 1]. N ewScore: Originally, each weak classiﬁer in the soft- cascade owns a score and ﬁnal score is the sum of all scores. Instead, we use the number of weak classiﬁer that the im- age patch passed positively as the new score. Therefore the upper limit of the new score is 2048 in our case. OverlapRerank: Given an image, multiple detections from multi-view detectors exist each with a score. For each detection, we ﬁrst calculate the number of overlapped de- tection it has (overlap threshold is 0.65) and then multiply score of each detection with a factor of its overlapping num- ber ranking1. Sumof Overlap: Instead of using overlapping as a multi- ply factor, here we use the sum of overlapped detections’ scores as the current detection’s new score. Detection merging: Apart from the Greedy∗ version of Non Maximum Suppression [6], we also use the detection combination introduced in [20]. It averages the locations of overlapped detections rather than suppresses them. Detection adjustment: As shown in Figure 4 (a), dif- ferent databases have different annotation styles of ground- truth. Speciﬁcally, AFLW has square annotations with nose located approximately at the center. AFW uses tight rect- angular bounding boxes as annotations with the eye-brow being the approximate upper bound. FDDB uses elliptical annotations bounding the whole head. As our detector is trained on AFLW and tested on AFW and FDDB, there ex- ist offsets in both detection position and scale. According to observations, the offsets vary as face pose changes. There- fore we adopt a view-speciﬁc detection adjustment to alle- viate the offsets. Note that the adjustment is constant for all images and faces in the same database, see Figure 4 (b) for 1A toy example: Det1: score: 10, nOverlap: 10; Det2: score: 9, nOver- lap: 20; Det3: score: 5, nOverlap: 5. After score re-ranking: Det1: score: 10 × 2 3 = 6.67; Det2: score: 9 × 3 3 = 9; Det3: score: 5 × 1 3 = 1.67. Reranking None Normalization NewScore OverlapRerank SumofOverlap AP (%) Merging AP (%) Greedy* NMS 91.7 Combination 93.4 91.7 93.5 92.9 95.0 93.7 Table 1. Comparison of different methods of score re-ranking and detection merging. details. Summary: According to experimental results (Table 1), OverlapRerank seems to be the best score re-ranking method. The underlying reason may be that true positives usually have many overlapped detections, while the false positives would only get a few responses. Therefore lever- aging this overlapping information in score re-ranking can reduce many false positives. However, in practice, overlap related methods and detection combination both cost much time to process, which is infeasible in a large majority of applications. We ﬁnally adopt N ormalization score re- ranking combined with Greedy∗ Non Maximum Suppres- sion for the sake of detection speed. 5. Experiments In this section, we compare our method with state-of- the-art methods on AFW and FDDB databases which con- tain challenging faces in the wild. In AFW, we com- pare with three commercial systems (Google Picasa, Face.com and Face++) and ﬁve academic methods (Shen et al. [21], Zhu et al. [27], DPM [7], multiHOG [27] and Kalal et al. [9]). In FDDB, we compare with one com- mercial system (Olaworks) and six academic methods (Yan et al. [23], Boosted Exemplar et al. [12], SURF multi- view [13], PEP-Adapt [11], XZJY [21] and Zhu et al. [27]) listed on FDDB results page1. 5.1. Evaluation on benchmark face database As shown in Figure 5, in AFW, our multi-scale detec- tor achieves an ap value of 96.8%, outperforming other academic methods by a large margin. When it comes to commercial systems, ours is better than Face.com and al- most equal to Face++ and Google Picasa. Note that most of our false positives on AFW database are faces that haven’t been annotated (small, seriously occluded or artiﬁ- cial faces like mask and cartoon character). When evaluated on FDDB database, we follow the eval- uation protocol in [8] and report the average discrete and continuous ROC of the ten subfolders. For equality, we ﬁx the number of false positives to 284 (equivalent to an av- erage of 1 False Positive Per Image) and compare the true 1http://vis-www.cs.umass.edu/fddb/results.html

Figure 5. Experimental results on AFW and FDDB database. Best viewed in color. positive rate. In discrete score where evaluation metric is the same as in AFW, our detector achieves 83.7%, which is a little better than Yan et al. [23]. Note that the ground- truth in FDDB are elliptical faces, therefore the evaluation metric of an overlap ratio bigger than 0.5 cannot reveal the true performance of the proposed detector well. When using continuous score which takes the overlap ratio as the score, our method gets 61.9% true positive rate at 1 FPPI for multi- scale version, surpassing other methods which output rect- angular detections by a notable margin (the Yan et al. de- tector outputs the same elliptical detections as the ground- truth, therefore having advantages with this metric). Our detector using single-scale features performs a little worse with the beneﬁt of faster detection speed. 5.2. Discussion Training efﬁciency: We implement the method with Pi- otr’s MATLAB toolbox [4] on a PC with Intel Core i7-3770 CPU and 16GB RAM. With 21, 328 positive images and 5, 771 negative images in total 6 views, the training pro- cess takes about 5.3 mins for a single-scale subview de- tector containing 2048 weak classiﬁers and 10.2 mins for multi-scale version. Note that we use much fewer training data than SURF multiview [13] whilst still outperforming their performance. Comparative results: When inspecting detections of the proposed face detector and other algorithms on the test- sets, some patterns can be found to explain why our detector outperforms others. One evident strength lies in detecting faces with extreme poses. Because we adopt multi-view detection and train each subview detector separately, our detector handles pose variations very well. Second is the outstanding illumination invariance of our detector, which is mainly owing to the extension of channel types to LUV color space and gradient-related channels. Detection speed: Due to the simple form of aggre- gate channel features and fast computation of feature pyra- mid [5], detection is quite efﬁcient. For full yaw pose face detection in VGA image, the proposed detector using single-scale features runs at 20 FPS on a single thread and 62 FPS if 4 threads are used. If only frontal faces are con- cerned, the detector runs at 34 FPS and 95 FPS after par- allelization. When it comes to the proposed detector using multi-scale features, the above four indices reduce to 15, 42, 21 and 55 FPS. Considering the large performance gain and similar speed, the proposed method can replace Viola-Jones detector for face detection in the wild. 6. Conclusion A novel feature representation called aggregate channel features possesses the merits of fast feature extraction and powerful representation capacity. In this paper, we success- fully apply the feature representation to face detection do- main through a deep investigation into the feature design, and propose a multi-scale version of feature which further enriches the representation capacity. Combined with our efforts into solving issues concerning multi-view detection, the proposed multi-view face detector shows state-of-the-art performance in both effectiveness and efﬁciency on faces in the wild. The proposed method appeals to real world appli- cation demands and has the potential to be embedded into low power devices. Acknowledgement This work was supported by the Chinese National Nat- ural Science Foundation Projects #61105023, #61103156, #61105037, #61203267, #61375037, National Science and Technology Support Program Project #2013BAK02B01, Chinese Academy of Sciences Project No. KGZD-EW-102- 2, and AuthenMetric R&D Funds. References [1] L. Bourdev and J. Brandt. Robust object detection via soft cascade. In CVPR, 2005. [2] S. C. Brubaker, J. Wu, J. Sun, M. D. Mullin, and J. M. Rehg. On the design of cascades of boosted ensembles for face detection. IJCV, 2008.

Figure 6. Results of comparative experiments in feature design. Best viewed in color. [3] N. Dalal and B. Triggs. Histograms of oriented gradients for human detection. In CVPR, 2005. [4] P. Doll´ar. Piotr’s image and video matlab toolbox (pmt). Soft- ware available at: http://vision.ucsd.edu/˜pdollar/ toolbox/doc/index.html, 2012. [5] P. Doll´ar, R. Appel, S. Belongie, and P. Perona. Fast feature pyramids for object detection. TPAMI, 2014. [6] P. Doll´ar, Z. Tu, P. Perona, and S. Belongie. Integral channel features. In BMVC, 2009. [7] P. F. Felzenszwalb, R. B. Girshick, D. McAllester, and D. Ramanan. Object detection with discriminatively trained part-based models. TPAMI, 2010. [8] V. Jain and E. Learned-Miller. Fddb: A benchmark for face detection in unconstrained settings. Technical report, 2010. [9] Z. Kalal, J. Matas, and K. Mikolajczyk. Weighted sampling for large- scale boosting. In BMVC, 2008. [10] M. Koestinger, P. Wohlhart, P. M. Roth, and H. Bischof. Annotated facial landmarks in the wild: A large-scale, real-world database for facial landmark localization. In First IEEE International Workshop on Benchmarking Facial Image Analysis Technologies, 2011. [11] H. Li, G. Hua, Z. Lin, J. Brandt, and J. Yang. Probabilistic elastic part model for unsupervised face detector adaptation. In ICCV, 2013. [12] H. Li, Z. Lin, J. Brandt, X. Shen, and G. Hua. Efﬁcient boosted exemplar-based face detection. In CVPR, 2014. [13] J. Li and Y. Zhang. Learning surf cascade for fast and accurate object detection. In CVPR, 2013. [14] S. Z. Li, L. Zhu, Z. Zhang, A. Blake, H. Zhang, and H. Shum. Sta- tistical learning of multi-view face detection. In ECCV. 2002. [15] S. Liao, A. K. Jain, and S. Z. Li. Unconstrained face detection. Tech- nical report, 2012. [16] M. Mathias, R. Benenson, M. Pedersoli, and L. Van Gool. Face detection without bells and whistles. In ECCV, 2014. [17] T. Mita, T. Kaneko, and O. Hori. Joint haar-like features for face detection. In ICCV, 2005. [18] M.-T. Pham and T.-J. Cham. Fast training and selection of haar fea- tures using statistics in boosting-based face detection. In ICCV, 2007. [19] M.-T. Pham, Y. Gao, V.-D. Hoang, and T.-J. Cham. Fast polygo- nal integration and its application in extending haar-like features to improve object detection. In CVPR, 2010. [20] P. Sermanet, D. Eigen, X. Zhang, M. Mathieu, R. Fergus, and Y. Le- Cun. Overfeat: Integrated recognition, localization and detection using convolutional networks. arXiv:1312.6229v4. [21] X. Shen, Z. Lin, J. Brandt, and Y. Wu. Detecting and aligning faces by image retrieval. In CVPR, 2013. [22] P. Viola and M. J. Jones. Robust real-time face detection. 2004. IJCV, [23] J. Yan, Z. Lei, L. Wen, and S. Z. Li. The fastest deformable part model for object detection. In CVPR, 2014. [24] M. D. Zeiler and R. Fergus. Stochastic pooling for regularization of deep convolutional neural networks. arXiv:1301.3557v1. [25] C. Zhang and P. A. Viola. Multiple-instance pruning for learning efﬁcient cascade detectors. In NIPS, 2007. [26] Q. Zhu, M.-C. Yeh, K.-T. Cheng, and S. Avidan. Fast human detec- tion using a cascade of histograms of oriented gradients. In CVPR, 2006. [27] X. Zhu and D. Ramanan. Face detection, pose estimation, and land- mark localization in the wild. In CVPR, 2012.

分享到：

赞收藏

资料库

ACF目标检测算法论文.pdf

相关推荐

人工智能

热门标签

最新资料