所需积分/C币:10 2019-03-23 21:56:14 6.5MB PDF

State-of-the-art object detection networks depend on region proposal algorithms to hypothesize object locations. Advances like SPPnet [1] and Fast R-CNN [2] have reduced the running time of these detection networks, exposing region proposal computation as a bottleneck. In this work, we introduce a R
classifier single, unified network for object detection(Figure 2) Using the recently popular terminology of neural Rol pooling networks with ' 31 mechanisms, the rPN module tells the Fast R-cnn module where to look In Section 3.1 we introduce the designs and properties proposals of the network for region proposal. In Section 3. 2 we evelop algorithms for training both modules with features shared Region Proposal Network 3.1 Region Proposal Networks feature maps A Region Proposal Network(RPN) takes an image (of any size) as input and outputs a set of rectangular object proposals, each with an objectness score.We model this process with a fully convolutional network [7l which we describe in this section Because our ulti- mate goal is to share computation with a Fast R-CNN object detection network [2], we assume that both nets share a common set of convolutional layers. In our ex periments, we investigate the Zeiler and Fergus model igure 2: Faster R-CNN is a single, unified network [32(ZF), which has 5 shareable convolutional layers Fi for object detection. The RPN module serves as the and the Simonyan and Zisserman model [3](VGG-16 attention of this unified network which has 13 shareable convolutional layers To generate region proposals, we slide a small network over the convolutional feature map output into a convolutional layer for detecting multiple class by the last shared convolutional layer. This small specific objects. The MultiBox methods [26], [27] gen network takes as input an m x n spatial window of erate region proposals from a network whose last the input convolutional feature map. Each sliding fully-connected layer simultaneously predicts mul- window is mapped to a lower-dimensional feature tiple class-agnostic boxes, generalizing the single.(256-d for ZF and 512-d for VGG, with ReLU [331 box"fashion of OverFeat. These class-agnostic boxes following). This feature is fed into two sibling fully are used as proposals for R-CNN 15]. The MultiBox connected layers-a box-regression layer(reg) and a proposal network is applied on a single image crop or box-classification layer(cls). We use n=3 in this multiple large image crops(e.g. 224x224), in contrast Paper, noting that the effective receptive field on the to our fully convolutional scheme. MultiBox does not put image is large(171 and 228 pixels for ZF and share features between the proposal and detection VGG, respectively). This mini-network is illustrated networks. We discuss overEat and multiBox in more at a single position in Figure 3 (left). Note that be depth later in context with our method. Concurrent cause the mini-network operates in a sliding- -window with our work, the DeepMask method [28] is devel- fashion, the fully-connected layers are shared across oped for learning segmentation proposals ll spatial locations. This architecture is naturally im- Shared computation of convolutions [91, [11, [291 Plemented with an n x n convolutional layer followed 71,[21 has been attracting increasing attention for ef. by two sibling I x I convolutional layers(for reg and ficient, yet accurate, visual recognition. The OverHeat cls, respectively) [9] computes convolutional features from an 3.1.1 Anchors image pyramid for classification, localization, and de At each sliding-window location we simultaneously tection. Adaptively-sized pooling(SPP)[1] on shared convolutional feature maps is developed for efficient predict multiple region proposals, where the number of maximum possible proposals for each location is [301 and semantic denoted as k. So the reglayer has 4k outputs encoding segmentation [29]. Fast R-Cnn [2] enables end-to-end g the coordinates of k boxes and the cls layer outputs detector training on shared convolutional features and shows compelling accuracy and speed 2k scores that estimate probability of object or not object for each proposal*. The k proposals are param eterized relative to k reference boxes which we call 3 FASTER R-CNN 3. Region" is a generic term and in this paper we only consider Our object detection system, called Faster R-CNN, 1S rectangular regions, as is common for many methods(e.8,[27, [4], composed of two modules. The first module is a deep [6)).Objectness"measures membership to a set of object classes fully convolutional network that proposes regions, US. background 4. For simplicity we implement the cls laver as a two-class and the second module is the Fast r-cnn detector [21 softmax layer. Alternatively, one may use logistic regression to that uses the proposed regions. The entire system is a produce k scores 2k scores 4k coordinates k anchor boxes cls layer reg layer :098 0.979 intermediate layer sliding window conv feature map Figure 3: Left: Region Proposal Network(RPN). Right: Example detections using rpn proposals on PASCAL VOC 2007 test. Our method detects objects in a wide range of scales and aspect ratios anchors. An anchor is centered at the sliding window Multi-Scale Anchors as regression References in question, and is associated with a scale and aspect Our design of anchors presents a novel scheme ratio(Figure 3, left). By default we use 3 scales and for addressing multiple scales(and aspect ratios). As 3 aspect ratios, yielding k=9 anchors at each sliding shown in Figure 1, there have been two popular ways position. For a convolutional feature map of a size for multi-scale predictions. The first way is based on w x H(typically 2, 400), there are W Hk anchors in image /feature pyramids, e.g. in DPM [8 and CNN total based methods [9],[1, [2]. The images are resized at Translation-Invariant Anchors multiple scales, and feature maps(HOG [8] or deer convolutional features [9 ,[1,[2])are computed for An important property of our approach is that it each scale(Figure 1(a)). This way is often useful but is translation invariant, both in terms of the anchors is time-consuming. The second way is to use sliding and the functions that compute proposals relative to windows of multiple scales(and /or aspect ratios )on the anchors. It one translates an object in an image, the feature maps. For example, in DPM [81, models the proposal should translate and the same function of different aspect ratios are trained separately using should be able to predict the proposal in either lo- different filter sizes(such as 5x7 and 7x5). If this way cation. This translation-invariant property is guaran- is is used to address multiple scales, it can be thought teed b y our method. As a comparison, the MultiBox of as a"pyramid of filters"(Figure 1(b). The second method [27] uses k-means to generate 800 anchors, way is usually adopted jointly with the first way [8 which are not translation invariant. So MultiBox does As a comparison our anchor-based method is built not guarantee that the same proposal is generated if on a pyramid of anchors, which is more cost-efficient an object is translated Our method classifies and regresses bounding boxes The translation-invariant property also reduces the with reference to anchor boxes of multiple scales and model size. Multi Box has a(4+1)X800-dimensional aspect ratios. It only relies on images and feature fully-connected output layer, whereas our method has maps of a single scale, and uses filters( sliding win a(4+2)x 9-dimensional convolutional output layer dows on the feature map)of a single size. We show by in the case of k=9 anchors. As a result, our output experiments the effects of this scheme for addressing ayer has28×10 parameters(512X(4+ 2)x9 multiple scales and sizes(Table 8) for VGG-16), two orders of magnitude fewer than Because of this multi-scale design based on anchors MultiBox's output layer that has 6. 1 x 10 parameters we can simply use the convolutional features com (1536x(4+1)x 800 for Google Net [34] in MultiBox puted on a single-scale image, as is also done by 271). If considering the feature projection layers, our the Fast R-cnn detector [2]. The design of multi proposal layers still have an order of magnitude fewer scale anchors is a key component for sharing parameters than MultiBox. We expect our method without extra cost for addressing scales to have less risk of overfitting on small datasets, like PASCAL VOC 3.1.2 Loss Function For training RPNs, we assign a binary class label 5. As is the case of FCNs [7, our network is translation invariant (of being an object or not) to each anchor. We as up to the network's total stride gn a positive label to two kinds of anchors: (i) the 6. Considering the feature projection layers, our proposal layers arameter count is3×3×512×512+512×6×9=2.4x106: anchor/ anchors with the highest Intersection-over- MultiBox's proposal layers Parameter count is 7 x 7 x(64+96+ Union(IoU)overlap with a ground-truth box, or (ii)an 64+64)×1536+1536×5×800=27×106 anchor that has an Iou overlap higher than 0.7 with any ground-truth box. Note that a single ground-truth be thought of as bounding-box regression from an box may assign positive labels to multiple anchors anchor box to a nearby ground-truth box Usually the second condition is sufficient to determine Nevertheless, our method achieves bounding-box the positive samples, but we still adopt the first regression by a different manner from previous rol- condition for the reason that in some rare cases the based(Region of Interest) methods [1],[2]. In [1] second condition may find no positive sample. We [2], bounding- box regression is performed on features assign a negative label to a non-positive anchor if its pooled from arbitrarily sized rols, and the regression loU ratio is lower than 0. 3 for all ground-truth boxes. weights are shared by all region sizes. In our formula Anchors that are neither positive nor negative do not tion the features used for regression are of the same contribute to the training objective spatial size (3 X 3)on the feature maps. To account With these definitions, we minimize an objective for varying sizes, a set of h bounding-box regressors function following the multi-task loss in Fast R-cnn are learned. Each regressor is responsible for one scale 2 Our loss function for an image is defined as and one aspect ratio and the k regressors do not share weights. As such, it is still possible to predict boxes of 1 L({p},{t})= ∑L(p,D;) various sizes even though the features are of a fixed size/ scale, thanks to the design of anchors reg t 3. 1. 3 Training RPNs The RPn can be trained end-to-end by back- Here, i is the index of an anchor in a mini-batch and propagation and stochastic gradient descent (SGD) pi is the predicted probability of anchor i being an object. The ground-truth label Pi is 1 if the anchor from [21 [35 We follow the"image-centric"sampling strategy to train this network. Fach mini-batch arises is positive, and is 0 if the anchor is negative. ti is a from a single image that contains many positive and vector representing the 4 parameterized coordinates of the predicted bounding box, and ti is that of the negative example anchors. It is possible to optimize for the loss functions of all anchors but this will ground-truth box associated with a positive anchor. bias towards negative samples as they are dominate The classification loss Lcls is log loss over two classes (object us. not object). For the regression loss, we use Instead, we randomly sample 256 anchors in an image Lreg(ti, t i )-=R(ti-ti)where R is the robust loss to compute the loss function of a mini-batch, where the sampled positive and negative anchors have a function (smooth I 1)defined in [2]. The term pi reg ratio of up to 1: 1. If there are fewer than 128 positive means the regression loss is activated only for positive samples in an image, we pad the mini-batch with anchors(p*= 1) and is disabled otherwise(p The outputs of the cls and reg layers consist of (pil negative and ti respectively. We randomly initialize all new layers by drawing weights from a zero-mean gaussian distribution with The two terms are normalized by Ncls and Nr reg standard deviation 0.01. All other layers (i. e, the parameter A. In our shared convolutional layers)are initialized by pre- current implementation(as in the released code), the training a model for ImageNet classification [361, as cls term in Eqn. (1) is normalized by the mini-batch is standard practice [5]. We tune all layers of the size (i.e., Ncls= 256)and the reg term is normalized ZF net, and conv 1 and up for the VGG net to by the number of anchor locations (i.e, Nreg N 2, 400) conserve memory [2]. We use a learning rate of 0.001 By default we set A=10, and thus both cls and for 60k mini-batches, and 0.0001 for the next 20k B reg terms are roughly equally weighted. We show mini-batches on the pascal voc dataset. We use a y experiments that the results are insensitive to the values of A in a wide range Table 9). We also momentum of 0.9 and a weight decay of 0.0005 [37] o note Our implementation uses Caffe [381 that the normalization as above is not required and could be simplified For bounding box regression, we adopt the param- 3.2 Sharing Features for RPN and Fast R-CNN eterizations of the 4 coordinates following 15 Thus far we have described how to train a network for region proposal generation, without considering a the region- based object detection Cnn that will utilize log(w/wa), th= log(h/h (2)these proposals. For the detection netw ork, we adopt =(m-)/l,t=(y-9)/ha log(w*/wa), th= log(h*/ha), learn a unified network composed of rpn and Fast R-CNN with shared There r, g, w, and h denote the box's center coordi- Both RPN and Fast R-CNN, trained independently, nates and its width and height. Variables a, aa, and will modify their convolutional layers in different .* are for the predicted box, anchor box, and ground- ways. We therefore need to develop a technique that truth box respectively(likewise for y, w, h). This can allows for sharing convolutional layers between the Table 1: the learned average proposal size for each anchor using the ZF net (numbers for s=600) anchor‖1282,2:11282,111282,12|262,212562,112562,12|512,2112,1512,1:2 proposal188×111113×11470×92416×229261×284174×332768×437499×501355×715 two networks, rather than learning two separate net- fix the shared convolutional layers and only fine-tune works. We discuss three ways for training networks the layers unique to RPN. Now the two networks with features shared share convolutional layers. Finally, keeping the shared (i)Alternating training. In this solution, we first train convolutional layers fixed, we fine-tune the unique RPN, and use the proposals to train Fast R-CNN. layers of Fast R-CNN. As such, both networks share The network tuned by Fast R-CNN is then used to the same convolutional layers and form a unified initialize RPN, and this process is iterated. This is the network. a similar alternating training can be run solution that is used in all experiments in this paper. for more iterations, but we have observed negligible (ii)Approximale joint training. In this solution, the mprovements RPN and Fast R-CNN networks are merged into one network during training as in Figure 2. In each SGD 3.3 Implementation Details Iteration nerates region propos- We train and test both region proposal and object ls which are treated just like fixed, pre-computed detection networks on images of a single scale [1], [2] proposals when training a FastR-cnn detector. The We re-scale le the images such that their shorter side backward propagation takes place as usual, where for Is s 600 pixels [2]. Multi-scale feature extraction the shared lavers th the backward propagated signals (using an image pyramid)may improve accuracy but from both the rpn loss and the fast r-cnn loss does not exhibit a good speed-accuracy trade-off [2] are combined. This solution is easy to implement. But On the re-scaled images, the total stride for both zF this solution ignores the derivative w.r.t. the proposal and vgg nets on the last convolutional layer is 16 boxes'coordinates that are also network responses, pixels, and thus is x10 pixels on a typical PASCal so is approximate. In our experiments, we have em- image before resizing( 500 x 375). Even such a large pirically found this solven uces close results, vet stride provides good results, though accuracy may be educes the training time by about 25-50% comparing further improved with a smaller stride with alternating training. This solver is included in For anchors, we use 3 scales with box areas of 1282 our released python code 2562, and 512- pixels, and 3 aspect ratios of 1: 1, 1: 2, (ii) Non-approximale joint training. As discussed and 2: 1.'These hyper-parameters are not carefully cho- above, the bounding boxes predicted by rPn are Sen for a particular dataset, and we provide ablation also functions of the input. The rol pooling layer experiments on their effects in the next section. As di [2]in Fast R-CNN accepts the convolutional features cussed, our solution does not need an image pyramid and also the predicted bounding boxes as input, so or filter pyramid to predict regions of multiple sca a theoretically valid backpropagation solver should saving considerable running time. Figure 3 (right) also involve gradients w.r. t the box coordinates. These shows the capability of our method for a wide range gradients are ignored in the above ap proximate joint of scales and aspect ratios. Table 1 shows the learned training. In a non-approximate joint training solution, average proposal size for each anchor using the ZF we need an roi pooling layer that is differentiable net. We note that our algorithm allows predictions w.r.t. the box coordinates. This is a nontrivial problem that are larger than the underlying receptive field and a solution can be given by an"Rol warping" layer Such predictions are not impossible--one may still as developed in [15], which is beyond the scope of this le Scope of this roughly infer the extent of an object if or e middle of the object is visible paper. 4-Step Alternating Training. In this paper, we ado o he anchor boxes that cross image boundaries need o be handled with care During trainin g, we igr a pragmatic 4-step training algorithm to learn shared all cross-boundary anchors so they do not contribute features via alternating optimization In the first step, to the loss. For a typical 1000 X 600 image,there we train the rpn as described in Section 3. 1.3. This will be roughly 20000( 60 X40 x 9 )anchors in network is initialized with an ImageNet-pre-trained total. With the cross-boundary anchors ignored, there model and fine-tuned end-to-end for the region pro- are about 6000 anchors per image for training. If the posal task. In the second step, we train a separate boundary-crossing outliers are not ignored in training detection network by Fast R-CNN using the proposals they introduce large, difficult to correct error terms in generated by the step-1 RPN. This detection net- the objective, and training does not converge. During work is also initialized by the ImageNet-pre-trained testing, however, we still apply the fully convolutional model. At this point the two networks do not share rpn to the entire image. This may generate cross convolutional layers. In the third step, we use the boundary proposal boxes, which we clip to the image detector network to initialize RPN training but we boundary Table 2: Detection results on PAsCaL voC 2007 test set(trained on vOC 2007 trainval). The detectors are Fast R-CNN with ZE, but using various proposal methods for training and testing train-time region proposals test-time region proposal method boxes method proposals mAP (%) 2000 587 EB 2000 EB 2000 58.6 RPN+ZF, shared 2000 RPN+ZF, shared 300 59.9 ablation experiments follow belou RPN+ZE, unshared 2000 RPN+ZE, unshared 300 58.7 2000 RPN+ZE 100 55.1 2000 RPN+ZF 568 RPN+ZE 1000 563 2000 RPN+ZF (no NMs) 6000 55.2 2000 RPN+ZF (no cls 100 44.6 2000 RPN+ZF (no cls) 514 2000 RPN+ZF (no cls) 1000 55.8 2000 RPN+ZF (no res 300 52.1 2000 RPN+ZF(no reg 1000 513 2000 RPN+VGG 300 592 Some rpn proposals highly overlap with each ToU. SS has an mAP of 58.7% and eB has an mAP other To reduce redundancy, we adopt non-maximum of 58.6% under the Fast R-cnn framework. RPN suppression(NMS)on the proposal regions based on with Fast R-CNN achieves competitive results, with their cls scores. We fix the IoU threshold for NMs an mAP of 59.9% while using up to 300 proposals at 0.7, which leaves us about 2000 proposal regions Using RPN yields a much faster detection system than per image. As we will show, NMS does not harm the using either SS or eB because of shared convolutional ultimate detection accuracy, but substantially reduces computations; the fewer proposals also reduce the the number of proposals. After NMS, we use the region-wise fully-connected layers cost (Table 5) top-N ranked proposal regions for detection. In the ablation experiments on rpn. to investigate the be following, we train Fast R-CNN using 2000 RPN pro- havior of RPNs as a proposal method, we conducted posals, but evaluate different numbers of proposals at several ablation studies. First we show the effect of test-time sharing convolutional layers between the rPn and Fast R-CNN detection network. To do this, we stop 4 EXPERIMENTS fter the second step in the 4-step training process 4.1 Experiments on PASCAL VOC Using separate networks reduces the result slightly to We comprehensively evaluate our method on the 58.7%(RPN+ZE, unshared, Table 2). We observe that PASCAL VOC 2007 detection benchmark [11]. This this is because in the third step when the detector dataset consists of about 5k trainval images and 5k tuned features are used to fine-tune the rpn, th test images over 20 object categories. We also provide proposal quality Is improve results on the pascal voc 2012 benchmark for a Next, we disentangle the RPN's influence on train few models. For the ImageNet pre-trained network ing the Fast r-cnn detection network.For th we use the"fast"version of ZF net [32] that has pose, we train a Fast r-CNn model by using the 5 convolutional layers and 3 fully-connected layers 2000 SS proposals and ZF net. We fix this detector ne public G-16 model" 3 th and evaluate the detection maP by changing the volutional lavers and 3 fully-connected lavers. We proposal regions used at test-time. In these ablation Primarily evaluate detection mean Average Precision experiments, the rpn does not share features with (mAP), because this is the actual metric for object the detector detection (rather than focusing on object proposal Replacing SS with 300 RPN proposals at test-time proxy metrics) leads to an maP of 56.8%. The loss in map is because Table 2 (top) Shows Fast R-cNN results when of the inconsistency between the training/testing pro- trained and tested using various region proposal posals. This result serves as the baseline for the fol- methods. These results use the zf net. For selective lowing comparisons Search(ss)[4 we generate about 2000 proposals b Somewhat surprisingly, the rPn still leads to a y the"fast"mode For Edge Boxes(EB)[6], we generate competitive result(55. 1%)when using the top-ranked the proposals by the default EB setting tuned for 0.7 8. For RPN, the number of proposals(e.g, 300)is the maximum number for an image. RPN may produce fewer proposals after NMS, and thus the average number of proposals is smaller Table 3: Detection results on PASCAL VOC 2007 test set. The detector is Fast R-CNN and VGG-16. Training data: 07: VOC 2007 trainval, 07+12" union set of voc 2007 trainval and voc 2012 trainval For rPn the train-time proposals for Fast R-CNN are 2000. t: this number was reported in [2]; using the repository provided by this paper, this result is higher (68.1) method proposals data AP(%) 2000 66.9 2000 07+12 70.0 RPN+VGG, unshared 300 RPN+ⅤGG, Shared 300 RPN+vGG, shared 07+12 73.2 RPN+vgg, shared 300COCO+07+12788 Table 4: Detection results on PASCAL VOC 2012 test set. The detector is Fast R-CNN and VGG-16. Training data: 07: VOC 2007 trainval, 07++12. union set of voc 2007 trainval+test and voc 2012 trainval. for Rpn,thetraintimeproposalsforFastR-cnnare2000.t: metho proposals data mAP (%o 2000 2 657 2000 07++12 RPN+ⅤGG, shared 300 12 67.0 RPN+VGG. shared 07++12 70.4 RPN+ⅤGG, shared 300 COCO+07++12 75.9 Table 5: Timing(ms)on a K40 GPU, except SS proposal is evaluated in a CPU. Region-wise" includes NMS, pooling fully-connected, and softmax layers. See our released code for the profiling of running time model system conv proposal region-wise total VGG SS+ Fast R-CNN 146 1510 174 1830 0.5 rPn Fast R-CN 10 ZF RPN Fast R-CNN 31 3 25 59 17 fps 100 proposals at test-time, indicating that the top-(using RPN+ZF) to 59. 2%(using RPN+VGG). This is a ranked rpn proposals are accurate. On the other promising result, because it suggests that the proposal extreme, using the top-ranked 6000 RPn proposals quality of RPN+VGG is better than that of RPN+ZH (without NMS) has a comparable mAp(55.2%), sug- Because proposals of RPn+ ZF are competitive with gesting NMS does not harm the detection mAP and Ss(both are 58.7% when consistently used for training may reduce false alarms and testing), we may expect RPN+VGG to be better Next, we separately investigate the roles of RPNs than SS. The following experiments justify this cls and reg outputs by turning off either of them pothesIs at test-time. When the cls layer is removed at test- time(thus no NMS ranking is used), we randomly Performance of VGG-16. Table 3 shows the results ample N proposals from the unscored regions. The of VGG-16 for both proposal and detection. Using mAP is nearly unchanged with N= 1000(55.8%), but RPN+VGG, the result is 68.5% for unshared features degrades considerably to 44.6% when /=100. This slightly higher than the Ss baseline. As shown above, shows that the cls scores account for the accuracy of this is because the proposals generated by RPN+VGG the highest ranked proposals are more accurate than SS. Unlike ss that is pre defined the rpn is actively trained and benefits from On the other hand, when the reg layer is removed better networks. for the feature-shared variant, the at test-time(so the proposals become anchor boxes), result is 69.9%-better than the strong SS baseline, yet the mAP drops to 52.1%. This suggests that the high- with nearly cost-free proposals. We further train the quality proposals are mainly due to the regressed box rpn and detection network on the union set of pas bounds. The anchor boxes, though having multiple cal voc 2007 trainval and 2012 trainval. The mAP scales and aspect ratios, are not sufficient for accurate is 73.2%. Figure 5 shows some results on the PASCAL detection VOC 2007 test set. On the pascal voc 2012 test set We also evaluate the effects of more powerful net-(Table 4), our method has an mAP of 70.4% trained works on the proposal quality of rpn alone. We use on the union set of voc 2007 trainval+test and VOC VGG-16 to train the rpn, and still use the above 20 12 trainval. Table 6 and table 7 show the detailed detector of SS+ZE. The mAP improves from 56.8% numbers Table 6: Results on pascal voc 2007 test set with Fast r-cnn detectors and VGG-16 For rpn. the train -time proposals for Fast R-Cnn are 2000. RPN* denotes the unsharing feature version method# hox data mAP ao hike bird boat bottle bus car cat chair cow table dog horse mike person plant sheep sofa train SS2000 07 669745783625323667378.282040772767.979679,273.069030:16547.2758658 2000 700770781693594383816 86742878868984782076.669931.870174.8 RPN" 30 07 68.574.177267753951.075.1 78.950.778061.179.181.972.275.937.271.462.5 23m424 RPn 300 6997080.670.157.349978280482.052275.367280379.875.076.339.168.367.381167.6 07+1273.276.57.0709655218318478652081965784884677576738873673.9830726 RN|300c00407+128843207768957881888963686370885987680.8235368047589667389 Table 7: Results on pascal voc 2012 test set with fast r-cnn detectors and vgg-16 for rpn the train -time proposals for Fast R-CNN are 2000 method# box data mAP areo bike bird boat bottle bus car cat chair cow table dog horse mbike person 2000 6578037766946937773968687741771.151.18607.879.8698 SS2000 07++12 82.378470.852.3 RPn300 376.471.048.4 82 77.871.689.344.273.055.087580.5 2.172.387342.273750.0868787 则Z35051m 58.9 RPN 07++12 70.484.979.8743 77.57.988.545.677.155.386981780.979640.172.660.981.261.5 RPN COCO+x+12|259874836768 81982091354982659.09.085584.784152278.965.585470.2 Table 8: Detection results of faster r-cnn on pas 3 scales and 3 aspect ratios(69.9% mAP in Table 8) CAL VOC 2007 test set using different settings of If using just one anchor at each position, the mAP anchors. The network is VGG-16. The training data drops by a considerable margin of 3-4%. The mAP is voc 2007 trainval. The default setting of using 3 is higher if using 3 scales (with 1 aspect ratio) or 3 scales and 3 aspect ratios (69.9%)is the same as that aspect ratios( with 1 scale), demonstrating that using n table 3 anchors of multiple sizes as the regression references settings anchor scales aspect ratios mAP( is an effective solution. Using just 3 scales with 1 1282 1:1 1 scale, i ratio 65.8 aspect ratio(69.8%)is as good as using 3 scales with 1:1 667 3 aspect ratios on this dataset, suggesting that scales 1 scale, 3 ratios {21,111:2H688 and aspect ratios are not disentangled dimensions for 256 21,1:1,1:2}67.9 the detection accuracy. But we still adopt these two 3 scales,, 1 ratio{1282,2562,512)1:1 698 3saes,3 ratios2,20.512){21,11,12}699 dimensions in our designs to keep our system flexible In Table 9 we compare different values of A in Equa- tion (1). By default we use A= 10 which makes the Table 9. Detection results of Faster r-cnn on pas- two terms in Equation (1) roughly equally weighted CAL VOC 2007 test set using different values of after normalization. Table 9 shows that our result is in equation(1). The network is VGG-16 The training impacted just marginally (by 1%)when d is within data is vOC 2007 trainval. The default setting of using a scale of about two orders of magnitude(1 to 100) =10(69.9%)is the same as that in Table 3 This demonstrates that the result is insensitive to a in 入 0.1 100 a wide range mAP(%)672689699691 Analysis of Recall-to-loU. Next we compute the recall of proposals at different lou ratios with ground truth boxes. It is noteworthy that the recall-to-loU In Table 5 we summarize the running time of the of the metric is just loosely 119l,[201 [21] related to the entire object detection system. Ss takes 1-2 seconds ultimate detection accuracy. It is more appropriate to depending on content (on average about 1.5s), and use this metric to diagnose the proposal method than to evaluate it Fast R-cnn with vGG-16 takes 320ms on 2000 SS proposals(or 223ms if using SVD on fully-connected In Figure 4, we show the results of using 300, 1000, layers [2). Our system with VGG-16 takes in total and 2000 proposals We compare with SS and EB, and 198ms for both proposal and detection. With the con the n proposals are the top- ranked ones based on volutional features shared, the rpn alone only takes the confidence generated by these methods. The plots 10ms computing the additional layers Our region- show that the rpn method behaves gracefully when ise computation is also lower, thanks to fewer pro the number of proposals drops from 2000 to 300. This posals(300 per image). Our system has a frame-rate explains why the rpn has a good ultimate detection of 17 fps with the Zf net maP when using as few as 300 proposals. As we analyzed before, this property is mainly attributed to Sensitivities to Hyper-parameters. In Table 8 we the cls term of the rPn. The recall of SS and eB drops investigate the settings of anchors. By default we use more quickly than rpn when the proposals are fewer 300 proposals 1000 proposals 2000 proposals 0.6 "+nss 02— RPN ZF RP 02— RPN ZF RPN VGG RPN VGG RPN VGG 0.6 8 lOU loU Figure 4: Recall us. loU overlap ratio on the PASCAL VOC 2007 test set Table 10: One-Stage Detection uS. Two-Stage Proposal Detection. Detection results are on the PASCaL vOC 2007 test set using the ZF model and Fast R-CNN. RPN uses unshared features proposals detector mAP(%) Two-Stage RPN+ ZF, unshared 300 Fast R-CNN ZF, 1 scale587 One-Stage dense, 3 scales, 3 aspect ratios 20000 Fast R-CNN+ZF, 1 38 One-Stage dense, 3 scales, 3 aspect ratios 20000 Fast R-CNN+ ZE, 5 scales 539 One-Stage Detection vs. Two-Stage Proposal + De- region proposals with sliding windows leads to 6% tection. The Over Feat paper [9] proposes a detection degradation in both papers. We also note that the one method that uses regressors and classifiers on sliding stage system is slower as it has considerably more windows over convolutional feature maps. OverFeat proposals to process is a one-stage, class-specific detection pipeline, and ours is a two-stage cascade consisting of class-agnostic pro posals and class-specific detections In OverFeat, the 4.2 Experiments on MS COcO region-wise features come from a sliding window of We present more results on the Microsoft COcO one aspect ratio over a scale pyramid. These features object detection dataset [12]. This dataset involves 80 are used to simultaneously determine the location and object categories. We experiment with the 80k images category of objects. In RPN, the features are from on the training set, 40k images on the validation set, square(3x3)sliding windows and predict proposals and 20k images on the test-dev set. We evaluate the relative to anchors with different scales and aspect mAP averaged for IOU E 0.5: 0.05: 0.95](COCO's ratios. Though both methods use sliding windowS, the standard metric, simply denoted as mAP@1.5, 95] region proposal task s only the first stage of Faster r- and mAP@0. 5(PASCAL VOC'S metric CNN-the downstream Fast r-cnn detector attends There are a few minor changes of our system made to the proposals to refine them In the second stage ot for this dataset. We train our models on an 8-GPU our cascade, the region-wise features are adaptively implementation and the effective mini-batch size be pooled [1], [2] from proposal boxes hat more faith- comes 8 for RPN (1 per GPU) and 16 for Fast R-CNN fully cover the features of the regions. We believe (2 per GPU). The rPn step and FastR-CNN step are these features lead to more accurate detections both trained for 240k iterations with a learning rate To compare the one-stage and two-stage systems, of 0.003 and then for 80k iterations with 0.0003.We we emulate the Over Feat system (and thus also circum- modify the learning rates(starting with 0.003 instead vent other differences of implementation details)by of 0.001)because the mini-batch size is changed. For one-stage Fast R-CNN. In this system, the proposals" the anchors we use 3 aspect ratios and scales are dense sliding windows of 3 scales(128, 256,512)(adding 642) rated by handling small and 3 aspect ratios (1: 1, 1: 2, 2: 1 ). Fast R-Cnn is objects on this dataset In addition, in our Fast R-CNN trained to predict class-specific scores and regress box step the negative samples are defined as those with locations from these sliding windows. Because the a maximum loU with ground truth in the interval of OverFeat system adopts an image pyramid, we also [0, 0.5), instead of [0. 1, 0.5 )used in [1], [2]. We note evaluate using convolutional features extracted from that in the SPPnet system [11, the negative samples 5 scales. We use those 5 scales as in [1, [21 in 0. 1, 0.5) are used for network fine-tuning but the Table 10 compares the two-stage system and two negative samples in[0, 0.5)are still visited in the SVm variants of the one-stage system. USing the ZF model, step with hard-negative mining But the Fast R-CNN the one-stage system has an mAP of 53.g%. This is system [2] abandons the SVM step, so the negative lower than the two-stage system(58.7%)by 4.8%0. samples in 0, 0. 1)are never visited Including these This experiment justifies the effectiveness of cascaded 0, 0. 1)samples improves mAP@0. 5 on the COCO region proposals and object detection. Similar obser- dataset for both Fast R-Cnn and Faster R-CNN sys vations are reported in [2] ,[39], where replacing ss tems(but the impact is negligible on PASCAL VOC)

试读 14P faster-rcnn

关注 私信 TA的资源

    faster-rcnn 10积分/C币 立即下载


    10积分/C币 立即下载 >