An Examination of Advances in Multistage Object Detection Techniques Utilizing Deep Learning

1.1 Background

Object detection methodologies, fundamental in a myriad of practical applications such as autonomous driving, video surveillance, and medical image analysis, rely heavily on precise object detection. The intricate nature of these methods facilitates higher accuracy, enabling objects to be identified and localized even within demanding and fluctuating environments. The progress within the field of object detection, therefore, holds transformative potential for the broader computer vision community, allowing for the development of more robust, precise, and efficient systems. This has far-reaching implications, enhancing the dependability and applicability of computer vision in everyday life.

1.png

Figure 1. Different groups of deep learning object detectors

Typically, an object detection framework encompasses two primary tasks: the classification task, which predicts the class labels of objects, and the localization task, which predicts the objects' locations within an image. The latter is often more challenging due to factors such as scale variation, occlusion, and viewpoint. Initial object detection methodologies were predominantly based on hand-crafted features and a linear classifier for object prediction within an image [1]. However, with the advent of deep convolutional neural networks (CNN) and its widespread applications in many fields [2, 3], object detection techniques rooted in deep learning have garnered substantial attention from the academic community.

As illustrated in Figure 1, deep learning-based object detection methodologies can be chiefly divided into four groups. Depending on the approach to refining object locations within an image, deep learning object detectors can be bifurcated into multi-stage and one-stage object detection techniques. The former first generates a series of proposal boxes indicative of object instances using a proposal generation network in the initial stage. These proposal features are then extracted and fed into subsequent networks for repeated object location refinement before final predictions are rendered. Conversely, one-stage object detection frameworks directly produce final predictions, bypassing object proposal generation. Given that multi-stage object detectors generate a sparse set of proposal boxes in the initial stage, they do not encounter the problem of class imbalance as one-stage object detectors do. Moreover, by refining object locations multiple times, the final bounding boxes generated by multi-stage object detectors are significantly improved compared to the predicted boxes generated by one-stage object detectors. However, multi-stage object detection frameworks typically possess a more complex structure than their one-stage counterparts, necessitating careful design to achieve optimal detection performance. Various benchmarks suggest that while multi-stage detection methods deliver superior detection accuracy, one-stage detection frameworks are characterized by simpler architecture and faster processing speed [4, 5]. It is noteworthy that two-stage object detection methods, due to their efficiency, have received most attention among multi-stage object detection techniques. This paper predominantly follows the design of two-stage object detection.

Figure 2. The structure of two-stage object detection framework

The proposal generation stage aims to produce a set of high-quality proposal boxes to represent object scales and locations from an input image. In this stage, an input image is fed into a deep backbone network to produce feature maps through different layers. Based on feature maps generated by the backbone, a proposal generation network is employed to generate a set of proposal boxes. Early deep CNN object detection methods usually exploit a single feature map generated by the backbone for producing proposal boxes. Recently, various techniques have been proposed to enhance semantic information of the backbone feature maps. In this paper, the feature extraction module includes the backbone network and extended modules. As an object can appear at any position and scale in an image, object representation methods are designed to produce initial guesses of object locations and shapes. Anchor box-based methods, which use a rectangular bounding box to represent an object, have led the fashion in the past few years. Recently, object detection methods based on point representations have emerged and attracted the attention of the research community due to their efficiency. As the final module in the proposal generation stage, proposal generation network uses object representations and input feature maps to generate high-quality proposal boxes. Current leading methods for proposal generation network are usually based on improving the region proposal network. In the following, we review recent optimization strategies in each module of the proposal generation stage.

3.1 Recent techniques in feature extraction

Deep learning-based object detection methods usually adopt a deep convolutional neural network pretrained on an image classification task as the backbone network to extract features from input images [30. 52]. In recent years, various deep architectures have been designed to improve classification performance. AlexNet [70] proposed an eight layers network with learnable parameters. ReLU activation and dropout layers are also adopted in the network to reduce the computational cost and prevent overfitting. VGG [18] proposed to use small convolution filters (i.e., 3×3 convolution filters) to increase the depth of the network. ResNets [19] introduced residual block which consists of a series of layers and a shortcut connection adding the input and output of the block. This design is very efficient to build a deeper network. ResNets is now still one of the most widely used backbone architectures for object detection models. Recently, various studies have focused on designing a light-weight deep architecture based on pointwise and depthwise convolutions for reducing computational complexity [71-73]. Li et al. [74] proved that using pretrained models designed for image classification tasks is not suitable for object detection tasks since object detection tasks not only need to classify the objects but also exactly localize the boundaries of the objects in image. The authors also proposed a deep network specially designed for object detection with high resolution feature maps to locate multi-scale objects. In general, object detection methods based on a single-level feature map (Figure 3) have a decline in detection accuracy due to scale variation of objects in natural scene images [23]. Moreover, the feature maps generated by the backbone usually contain high-level semantic information in deep layers and low-level semantic information in shallow layers. As a result, object detection methods usually apply the detection network on deep feature layers. However, due to low resolution in deep layers, the structure of objects may be destroyed, especially for small objects. This may compromise the detection performance of the detection network. In recent years, various methods have been proposed to improve the detection performance based on feature maps generated by the backbone network. According to the way of exploiting the backbone feature maps for detecting objects, these methods can be roughly divided into two groups: methods based on multi-layer backbone and methods based on feature pyramid. Detection methods based on multi-layer backbone adopt different feature maps at different layers of the backbone to predict objects while detection methods based on feature pyramid construct a high-level semantic feature pyramid to produce final predictions. It should be noted that since we focus on the two-stage object detection design, some techniques specially proposed for one-stage detection framework, such as deconvolution module [75], are not reviewed in this paper.

3.png

Figure 3. Object detection based on a single-level feature map [30, 52]

3.1.1 Detection methods based on multi-layer backbone

Multi-layer object detection methods propose to directly predict objects from different feature maps at different layers of the backbone network (Figure 4) [20-22]. These methods exploit features at different layers independently to overcome the scale variation problem since different feature layers encode information of objects at different scales. Multi-layer backbone object detection methods have received less attention in recent years. Most multi-layer backbone object detectors focus on producing high-level semantic feature maps by an optimization backbone network. MS-CNN [20] is the first two-stage object detection framework that applies the proposal generation network at different output layers of the backbone network. To bridge the scale gap between input image and convolutional layers, deconvolution layers are employed to increase the resolution of feature maps. However, RoI pooling scheme is still applied on a single feature map. This leads to misalignment between proposal boxes and object features, which compromises the detection performance. Li et al. [21] designed a scale-aware network which incorporates two different subnetworks into a unified architecture to deal with the scale variation of objects in image. Each subnetwork first uses convolution layers to further extract scale-specific feature maps based on feature maps generated by the backbone and then generates scale-specific detection results. Yang et al. [22] proposed a new approach to perform multi-scale object detection which first examines the size of each object proposal and then pools the features from a corresponding feature map. Three different layers in the VGG16 backbone are adopted to pool object features. To ensure the features are discriminative enough, a scale-dependent pooling scheme was introduced to provides strong supervision to enforce more discriminative convolutional filters.

4.png

Figure 4. Object detection based on multi-layer features [20-22]

3.1.2 Detection methods based on feature pyramid

In deep convolutional neural networks, shallow layers generally contain weak semantic representations and rich geometric information while deep layers have rich semantic representations but weak geometric information. Since multi-layer detection methods adopt features at different layers independently without integrating different feature levels, they do not achieve good detection results. Recently, with the advent of feature pyramid network [23], many methods based on feature pyramid backbone have been proposed to generate high-level semantic feature maps to improve the detection performance. Instead of exploiting different feature layers independently, feature pyramid detection methods first build a high-level semantic feature pyramid and then adopt each feature map in the pyramid to produce predictions. FPN [23] is the first framework that combines low-level features with high-level features via a top-down pathway and lateral connections to generate high-level semantic feature maps at all scales of the backbone. Figure 5 (a) illustrates the structure of FPN. The top-down pathway includes upsampling operations to upsample feature maps from higher pyramid levels to increase their spatial resolution and lateral connection operations to enhance high-level features with features from the bottom-up pathway. In addition, a 3×3 convolution is applied on each merged feature map to generate final feature maps to reduce the aliasing effect of upsampling operations. By using FPN as the feature extraction network, Faster R-CNN achieves an AP of 33.9 on the COCO minival set. Compared with the original Faster R-CNN based on VGG16, Faster R-CNN with FPN improves AP by 2.3 points. FPN quickly becomes an essential component in the feature extraction network in modern object detectors [6, 76, 77]. Despite the benefits, the straightforward integration in FPN makes it suffer from feature-level imbalance. To be more specific, feature fusion by simple operation ignores the semantic gap between different feature maps at different levels. In addition, semantic feature maps suffer from information loss due to 1×1 convolutional layers in lateral connections, especially with high-level layers. Various techniques have been proposed to overcome the shortcomings of FPN. In PANet [24], the authors suggested that low-level features are helpful in improving the localization capability of the detection network. Based on the idea, they designed a bottom-up path augmentation (Figure 5 (b)) to enhance entire feature pyramid with accurate localization signals existing in low-level features and shorten information path. The bottom-up path augmentation starts from the lowest feature map generated by FPN and gradually reaches the highest feature map. In each block, a 3×3 convolutional layer with stride 2 is first used to reduce the spatial size. Each feature map of FPN is then fused with the down-sampled feature map by lateral connection. Finally, the fused feature map is fed into a 3×3 convolutional layer to generate intermediate map for following subnetworks. Libra R-CNN [25] proposed to enhance multi-level feature maps generated by the FPN backbone by using integrated balanced semantic feature map (Figure 5 (c)). In this method, all feature maps from the backbone are first resized to an intermediate size. The balanced semantic feature map is obtained by averaging all resized feature maps from different layers. The balanced feature map is then refined to be more discriminative. Finally, the output feature maps are generated by rescaling the refined feature map. Overall, this technique will further strengthen the original features and reduce the imbalance at feature level which limits the overall detection performance. On the other hand, based on the observation that feature map at the highest level of FPN suffers from information lost due to channel reduction of the feature map, AugFPN [26] proposed a residual feature augmentation subnet to enhance feature representations of the highest-level backbone feature map (i.e., P5) (Figure 5 (d)). In the residual feature augmentation subnet, ratio-invariant adaptive pooling based on PSP [78] is first attached to the last feature map of the backbone (i.e., C5) to generate different context feature maps with different scales. These context feature maps are then rescaled and combined by a novel adaptive spatial fusion module to produce the final context feature map, which is later fused with the last reduced feature map (i.e., M5) to generate enhanced feature map. The authors also introduced a consistent supervision scheme which applies the same supervision signals on the feature maps after lateral connection to narrow the semantic gaps between feature maps. Another approach is to alleviate unequal contribution of input features in multi-scale feature fusion. Tan et al. [27] introduced an effective weighted bi-directional feature pyramid network (BiFPN) (Figure 5 (e)) with learnable weights to learn the contribution of different inputs at different resolutions. Based on feature pyramid generated by the EfficientNet backbone [79], BiFPN removes feature maps in top-down path with only one original input feature map (i.e., P3, P7). In the bottom-up path, original input feature maps are added to corresponding output feature maps if they are at the same pyramid level. Each bidirectional group, which includes top-down and bottom-up path, is treated as one feature layer so bidirectional group is repeated multiple times to add more high-level feature fusion to the feature extraction network. BiFPN also introduced fast and efficient normalized fusion algorithm to add an additional weight for each input when fusing feature maps at different layers. Different from BiFPN, A²-FPN [28] designed an attention aggregation pipeline based on FPN to refine multi-scale feature fusion through attention-guided feature aggregation (Figure 5(f)). The proposed pipeline extracts and fuses feature pyramid progressively through three modules. Multi-level global context module is first designed to replace 1×1 convolution in FPN to mitigate information loss due to channel reductions. Then, to reduce the effects of semantic gap between different layers in feature fusion stage, global attention module (GAM) is designed based on CARAFE [80] to improve the semantic consistency of adjacent feature maps before merging. Finally, global attention content-aware pooling (GACAP) is introduced in the bottom-up path to aggregate more discriminative information for output feature maps. Another approach is proposed by Qiao et al. [29]. The authors introduced Recursive Feature Pyramid (RFP) based on recursive convolutional network (Figure 5(g)) which adds feedback connections from output features into bottom-up path of the FPN network. The RFP modifies the first block of each stage of the ResNet backbone to update bottom-up patch with RFP features via atrous spatial pyramid pooling (ASSP) [81]. A feature fusion module is designed to update output feature maps of the RFP. In addition, all standard 3×3 convolutional layers in the bottom-up path of FPN are replaced by switchable atrous convolution to effectively enlarge the field-of-view of filters.

3.1.3 Comparison and analysis

Current state-of-the-art detectors utilize feature pyramid networks due to their robust performance and balance between speed and accuracy. Compared to Faster R-CNN with a ResNet-101 backbone, Faster R-CNN with FPN shows a slight improvement of 0.5 points in Average Precision (AP) on the COCO test-dev benchmark. While this increase is minor, it's achieved with no significant increase in computational cost. However, Libra R-CNN outperforms FPN, achieving 2.5 points higher AP on the same dataset. This showcases the enhanced accuracy of Libra R-CNN, but it should be noted that this improvement might come at the expense of increased complexity in model architecture and potentially higher computational cost. AugFPN, with its residual feature augmentation module, furthers the improvement by 1 AP point on the COCO validation set when compared to FPN, indicating its robustness in detection tasks. However, the model might involve more complexity due to the augmentation module, adding to the computational burden. BiFPN in EfficientDet offers significant enhancements, improving AP by 4 points compared to FPN-based ResNet-50 on the COCO validation set. Additionally, BiFPN requires 3× fewer parameters and 4× fewer FLOPs compared to FPN, presenting a major leap in computational efficiency. This positions it as a powerful choice for object detection tasks, particularly when computational resources are limited. By refining multi-scale feature fusion through three new modules, A²-FPN improves the detection performance by 2.4 points AP compared with the feature extraction network in PANet on the COCO validation set. However, the increase in computational complexity with A²-FPN might limit its applicability in resource-constrained settings, hence a lighter variant A²-FPN-Lite is introduced as a compromise between speed and accuracy. DetectoRS and its RPF network show considerable promise, especially in localizing occluded objects in images, improving AP by 4.2 points over FPN on the COCO validation set. The ability to exploit nearby context information to locate occluded parts of an object gives RPF a distinct advantage. However, the specific mechanism to achieve this might increase the model's complexity and computational demand. Meanwhile, multi-layer object detection methods, though less common, demonstrate potential. For instance, the method by Yang et al. [22] reported a significant relative improvement of 9% over Faster R-CNN on the KITTI test set. These methods, however, are deeply reliant on the backbone architecture for their performance, which might restrict their adaptability. It's worth noting that some feature extraction networks, originally introduced for one-stage object detection frameworks for efficiency, can also be effectively applied to multi-stage frameworks. This demonstrates a level of flexibility in deploying these networks, although the specific pros and cons will largely depend on the context of the application and the specific framework in use.

3.1.4 Open issues

Object detection methods that employ multi-level backbone features have received limited attention in recent years due to the inherent inefficiencies in independently predicting objects at different layers. Most state-of-the-art object detectors currently utilize a feature pyramid approach, prioritizing enhancement of the representation capacity of feature pyramids so that each feature level contains strong semantic information. However, merging feature maps of varying resolutions and representations from these pyramids poses a significant challenge due to potential loss of semantic information. Recent developments, such as A2-FPN [28] and DetectoRS [29], have addressed this issue by designing additional networks to mitigate information loss and tackle semantic gap problems. These methods have demonstrated promising results, but their increased complexity-with the addition of extra networks-invariably leads to more parameters and hyperparameters. This, in turn, introduces larger computational costs and requires more manual adjustments, an aspect that may not be feasible in all scenarios, particularly in resource-constrained environments. Another approach to improving the effectiveness of feature pyramid utilization is the use of Neural Architecture Search (NAS) [82] to identify the optimal FPN structure [83]. While promising, this approach is also computationally intensive, especially during the search process. To address these issues, future work could explore methods that balance accuracy and computational efficiency. For instance, lightweight neural architectures could be developed to minimize the parameters while maintaining detection performance. Techniques such as network pruning and quantization could also be used to optimize existing architectures. Furthermore, automated hyperparameter tuning algorithms, like Bayesian optimization, could mitigate the need for extensive manual tweaks. Ultimately, the objective is to develop an efficient method that fully exploits each feature level to generate increasingly powerful feature levels without exacerbating computational demands.

3.2 Recent techniques in object representations

According to the way of representing objects in the proposal generation stage, existing object detection methods can be categorized into two groups: anchor box-based methods and point-based methods. Anchor box-based methods represent each object as a rectangular bounding box on feature maps while point-based methods employ points on feature maps to depict objects. Early two-stage object detectors usually employ a rectangular bounding box representation scheme in the proposal generation stage to produce proposal boxes. Recently, with the emergence of FPN [23] and Focal loss [6], point representations have attracted increasing attention from researchers due to their efficiency. It should be noted that although point representations are usually employed in one-stage object detection pipelines, they can be modified to adapt to two-stage object detection pipelines to improve the overall detection performance [37, 38, 44, 50, 51, 84].

3.2.1 Rectangular bounding box representations

Early two-stage object detectors are usually based on rectangle bounding boxes (i.e., anchor boxes, proposal boxes, and final detection boxes) to represent objects at different recognition stages of the detection process [30, 55, 85]. A bounding box B={x, y, w, h} is a 4-d representation containing the spatial location of an object in image, with (x, y) representing the center points and (w, h) representing the width and height of object in image. The rectangular bounding box is the dominant type of object representation in early deep learning-based object detection frameworks, especially in two-stage detection methods, due to its convenience and efficiency [44]. In the proposal generation stage of two-stage object detection pipelines, input rectangular bounding boxes are usually called anchor boxes which are hypothesized to represent objects at different scales and aspect ratios. For each anchor box, features are extracted as object representations which are then used to generate proposal boxes by a proposal generation network. In general, a set of anchor boxes are generated at each location on the feature map by predefined scheme [7, 23, 30] or learnt scheme [31-33]. Anchor box generation based on predefined scheme is first introduced in Faster R-CNN [30], where an anchor box is centered at the sliding window and associated with a scale and aspect ratio. There are three predefined scales (i.e., 128², 256², 512²) and three predefined aspect ratios (i.e., 1:1, 1:2, 2:1), resulting nine anchor boxes at each position on the feature map in Faster R-CNN. In the study [23], RPN employs different scales at different feature layers to define anchor boxes. For anchor shape, three predefined aspect ratios are used, yielding 15 anchor boxes at each location over the feature pyramid. Recently, ThunderNet [7] has used a single scale for each feature layer and five aspect ratios (i.e., 1:2, 3:4, 1:1, 4:3, 2:1) to generate anchor boxes.

In the predefined anchoring scheme, the number of anchor boxes and the size of each box need to be designed carefully. To be more specific, too few anchor boxes may be insufficient to cover a large range of objects in various sizes and ratios, thus hindering the detection accuracy of the detector. In contrast, too many anchor boxes generated at the beginning require more parameters, which may lead to overfitting and significant computational cost due to a large number of false candidates. In addition, the anchor box shapes have to be manually tweaked to improve detection accuracy of the detector on specific domains. For example, since texts may have large aspect ratios compared with generic objects, Liao et al. [86] used seven predefined aspect ratios (i.e., 5:1, 3:1, 2:1, 1:1, 1:2, 1:3, 1:5) for scene text detection. In the study [87], the authors used one aspect ratio (i.e., width/height=5) for license plate detection since license plates are usually rectangular in shape. Wang et al. [88] employed single aspect ratio of 0.41 for pedestrian detection as this value is the average aspect ratio of pedestrians. To mitigate the issues of predefined anchoring scheme, learnt anchoring scheme [31-33] proposes to generate learnable anchor boxes, where anchor shape is learnt during the training process to generate high-quality anchor boxes to boost the detection performance. MetaAnchor [31] introduced an anchor function generator which maps any box prior to corresponding anchor function. The anchor function generator is formulated as a simple two-layer network and computed from customized prior boxes. Based on neural network weight prediction mechanism, anchor function generator could be implemented and embedded into existing object detection methods for joint optimization. The proposed mechanism is shown to be robust to anchor settings as it may cover various kinds of object boxes with any shape. However, the method shows minor improvements for two-stage anchor box-based object detection pipeline and requires customized prior boxes to be chosen by handcraft. In addition, it introduces an extra network for predicting weights, which leads to increased parameters and computational cost. Similar to the MetaAnchor scheme, Zhong et al. [32] introduced an anchor optimization scheme in which anchor shape is automatically learned during the training process. A localization loss is introduced which is to minimize the error between the ground truth box and the predicted offset relative to the anchor box. The error is then backpropagated to the anchor shapes and other parameters in the whole network to automatically learn the anchor box. The anchor shape is warmed up by soft assignment and online clustering scheme. Since the anchor shape is learned during training, the optimization scheme is more suitable for the specific data and network structure. In addition, the anchor shape initialization needs to be designed carefully to achieve the best detection performance. Recently, Sparse R-CNN [33] defined a fixed small set of learned candidate boxes to represent objects. Each candidate box includes a learnable proposal box and a learnable proposal feature. Proposal box is a 4-d representation containing initial center coordinates, height and width of object while proposal feature is a high-dimension vector which encodes rich information of object. The parameters of proposal boxes and features will be updated with the backpropagation algorithm during the training process and optimized together with other parameters in the whole network.

Although rectangular bounding box representations are easy to implement and facilitate computational process, they provide only coarse localization of objects. As a result, the feature extraction process in the object prediction stage will produce coarse representations of objects. The coarse extracted features may be heavily influenced by background information and foreground regions that contain little semantic information, which leads to inferior detection performance in object detection.

3.2.2 Point representations

To tackle the shortcomings of rectangular bounding box representations, various methods have been proposed to develop more efficient object representations. Recent state-of-the-art object detection methods directly employ points on the feature maps to represent objects and form bounding boxes based on proposal points. For simplicity, we call these methods as point representations. Point representations can be grouped into two groups: anchor-point representations and key-point representations. Anchor-point representations, including DenseBox [34], FSAF [35], Guided Anchoring [36], FCOS [37], FoveaBox [38], and SPAD [39], employ each pixel on a feature map as object representation. The proposal generation network first classifies each pixel into foreground or background class, and then directly regresses the distances from the foreground point to the four sides of the ground truth bounding box to generate final prediction. DenseBox [34] (Figure 6 (a)) proposed object detection based on FCNs [89]. DenseBox directly predicts a 4-d vector and a confident score of being an object at each location on a feature map. The 4-d vector represents the relative offsets from the top-left and bottom-right boundaries of the target bounding box to pixel location. To solve shape variations of object, DenseBox employs image pyramid by cropping and resizing input images to different sizes. Different from DenseBox, Zhu et al. [35] introduced a novel feature selective anchor-free module (FSAF) (Figure 6 (b)) which takes locations on the feature pyramid as inputs and directly feeds these locations into two convolutional branches: a classification branch for predicting K class scores for each location and a regression branch for producing 4 offsets encoding the distances between the current pixel location and the top, left, bottom, and right boundaries of the target bounding box. The FSAF module can be inserted into an anchor box-based detector with a feature pyramid backbone to help learning objects which are hard to be modeled by anchor-based mechanism. Another approach, Guided Anchoring [36] (Figure 6(c)), predicts each location on the feature map by using probability maps generated by an anchor location prediction network. At each active location where the center of objects is likely to exist, anchor shape prediction network is designed to predict the best shape for corresponding object. Recently, FCOS [6] directly classified each point on a feature map and regresses the target bounding box for positive points, thus eliminating complicated problems related to anchor bounding boxes. A point is considered as a positive point if it falls into any ground truth box and the class label of the point is the class label of the ground truth box (Figure 6 (d)). If a location falls into multiple ground truth bounding boxes, FCOS based on the multi-level features generated by FPN to assign an appropriate bounding box. In addition, FCOS introduced a novel center-ness branch with only one convolution layer to generate a centerness score for a positive point based on the distance from the point to the center of the corresponding object. This score is then used to eliminate low-quality bounding boxes. FCOS achieved state-of-the-art detection performance with much less design complexity and computational cost, which encourages researchers to follow the anchor-free mechanism in designing object detection pipeline. Similar to FCOS, FoveaBox [38] directly predicts the object score and the corresponding boundaries for each spatial location on the feature pyramid. FoveaBox uses fovea area, which is generated by shrinking ground truth box based on a shrunk factor, to define positive anchor points (Figure 6 (e)). In SAPD [39], the authors proved that the training strategy used in most anchor-point detectors, which treats anchor points independently, may lead to compromise the detection performance due to different contributions of anchor points to the network loss according to their spatial location. Based on this, they proposed a soft-weighted scheme which adds an attention weight for each anchor point based on its geometrical relation with the instance boundaries. The soft-weighted scheme forces the network to focus more on positive points near the center of object instance.

Alternatively, key-point representations, including CornerNet [40], CenterNet [41], ExtremeNet [42], Objects as Points [43], RepPoints [44], VarifocalNet [45], use keypoints such as center points, corners points, or extreme points to represent an object. These methods first predict the locations of keypoints of the bounding box and then group those key points to form a bounding box. CornerNet [40] proposed to represent an object as a pair of corner keypoints covering the object (i.e., the top-left corner and bottom-right corner of the bounding box) (Figure 7 (a)). The authors also introduced a novel corner pooling layer to better localize each type of corner of bounding boxes by employing prior knowledge. To group corner points that belong to the same object and produce a bounding box, the network first predicts an embedding vector for each detected corner and then groups corners based on the distances between their embeddings. Based on CornerNet, CenterNet [41] adds one extra keypoint (i.e., a center keypoint) to represent an object (Figure 7 (b)). While conner keypoints are used to generate bounding boxes, the center keypoint is employed to filter out incorrect bounding boxes. A center pooling layer is designed in CenterNet to capture rich semantic information within the center of the bounding box, thus improving the detection of the center keypoint. In addition, a cascade corner pooling layer is proposed to add center information to boundary information to improve conner localization capability of the network. Instead of using conner points to represent objects, ExtremeNet [42] defines four extreme points (i.e., top-most, left-most, bottom-most, right-most) and a center point to represent an object (Figure 7 (c)). Unlike conner points, extreme points usually lie on object, thus containing strong semantic information. This facilitates the extreme point detection performance. ExtremeNet also introduced a center grouping algorithm which analyses the geometric structure of extreme points and their center to group extreme points and produce bounding box. To eliminate the grouping stage in key-point detectors, Zhou et al. [43] represented objects by a single point at the center of bounding box (Figure 7 (d)). Peaks in the heatmap generated by the keypoint estimation network correspond to object center point. The keypoint prediction network predicts the height and width of bounding boxes covering objects based on the keypoint values of the center point. Different from corner points and extreme points, RepPoints [44] defined a set of adaptive sample points (e.g., 9 points) with the deformable convolution [54] to represent an object (Figure 7 (e)). RepPoints starts from the center point and produces other points via regressing offset values over the center point. For refinement object bounding box, RepPoints produces a 2-d regression vector (i.e., (∆x; ∆y)) instead of 4-d regression vector in most modern object detectors to alleviate the problem of scale differences among the bounding box regression parameters. For forming bounding boxes, RepPoints designed a converting function which produces bounding box based on adaptive sample points. Recently, VarifocalNet [45] defined a star-shaped bounding box with nine fixed sampling points to represent an object (Figure 7 (f)). The star-shaped bounding box representation can capture the geometric relations between a bounding box and its nearby contextual information. This facilitates the regression process which encodes the misalignment between the predicted box and the ground truth box.

Figure 6. Object detection based on anchor-point representations: (a) Densebox; (b) FSAF; (c) Guided Anchoring; (d) FCOS; (e) FoveaBox

7.png

Figure 7. Key-points representation in object detection: (a) CornerNet; (b) CenterNet; (c) ExtremeNet; (d) Objects as points; (e) RepPoints; (f) VarifocalNet

8.png

Figure 8. The architecture of RPDet [44]

Anchor-point and key-point representations are usually used in one-stage object detection frameworks due to their efficiency. However, they can be modified to adapt to two-stage object detection frameworks to improve the overall detection performance [33, 37, 38, 44, 50, 51, 84]. For example, FCOS [37] proposed to replace the anchor box scheme with the anchor point scheme in RPN. The results showed that RPN with the anchor-point scheme boosts the localization capability of the network by a large margin. In RPDet [44], a two-stage object detection approach (Figure 8), a set of refined points representing objects are produced by a localization subnet in the first stage. Proposal boxes are generated based on refined points by using a converting layer. These refined points and proposal boxes are then fed into a classification subnet at the second stage. Compared with baseline detectors based on bounding box representations, RPDet significantly improves the detection performance on the same dataset.

3.2.3 Comparison and analysis

Object detection methods utilizing rectangular bounding box representations have primarily pivoted towards learnt anchoring schemes in recent times. For instance, in the study [36], the Faster R-CNN improved detection performance by 2.7 points AP on the COCO test-dev set, by replacing the predefined anchor scheme with a guided one to automatically learn anchor shape and location. The benefit of this approach is the production of high-quality proposal boxes, increasing the efficiency of the training process. Yet, it also introduces more complexity to the model. In a similar vein, Zhong et al. [32] introduced a learning scheme to learn anchor shapes during the training process. The result was consistent improvements across different datasets and architectures with negligible extra training and inference costs. Notably, this learning scheme proved more robust with one-stage baselines compared to two-stage ones, due to the efficiency reduction caused by the second stage in two-stage baselines. Meanwhile, Sparse R-CNN [33], utilizing learnable proposal boxes, significantly outperformed Faster R-CNN based on predefined anchor boxes. However, its superior performance comes with the cost of a more complex network structure and possibly higher computational load.

Recently, the efficiency of point representations, aided by FPN and Focal loss, has been a focal point of research. Anchor-point representations, particularly recent methods [37-39], have achieved substantial improvements over anchor box-based methods. Notably, FCOS [38] outperformed both RetinaNet and Faster R-CNN by substantial margins on the test-dev split of the MS-COCO benchmark. Yet, these methods may potentially be more complex and resource-intensive than those using box representations. In the domain of key-point representations, CornerNet [40], an innovative one-stage approach, achieved competitive detection performance compared with several anchor box-based two-stage methods. The ExtremeNet [42] improved on this by combining extreme points and center point, achieving a modest increase in performance. Recently, VarifocalNet [45], which defines nine fixed sampling points to represent an object, has surpassed almost all state-of-the-art detectors. This result is promising for future research, but again, the complexity and computational cost may be a drawback. In comparison, anchor-point methods have simpler network architectures, which leads to faster training and inference speeds. However, key-point methods can offer superior accuracy by encoding geometric relations between an object and its nearby contextual information, highlighting a trade-off between complexity, computational cost, and detection performance in the field of object detection methods.

3.2.4 Open issues

Firstly, the efficiency of point representation strategies has led to their adoption in multi-stage object detection pipelines in several recent studies [33, 37, 38, 44, 50, 51, 84]. However, these strategies have not been explored as extensively in multi-stage object detectors as in one-stage object detectors. For example, FCOS [37] used the anchor-point scheme to replace the anchor-box scheme in RPN, which improved the localization capability of RPN. Yet, the effects of the anchor-point scheme on the second stage have not been thoroughly investigated. Potential solutions to address this issue could be conducting further research specifically focusing on the implications of the anchor-point scheme on the second stage and modifying the scheme to better suit the second stage based on the findings.

Secondly, the fine localization capabilities of point representations have drawn increasing attention. In the study [70], the authors identified the main difference between anchor box-based and point-based detectors as the sampling strategy (Section 5.3). Following this, they proposed an adaptive training sample selection strategy that automatically divides positive and negative samples based on statistical characteristics of the object. This sampling strategy allowed RetinaNet [6] with an anchor box representation scheme to achieve comparable detection performance with FCOS [37] with an anchor-point representation scheme. However, this strategy was proposed for one-stage object detectors, leaving the same issue for multi-stage object detectors unexplored. In response to this, future studies could aim to adapt this strategy to multi-stage object detectors and analyze its effect on detection performance.

Thirdly, despite key-point detectors being time-consuming and complex models due to their grouping algorithm, they have the advantage of superior detection accuracy compared to anchor-point detectors. Several studies proposed combining anchor-point and key-point representations to encode multiple levels of objects features and avoid the time-consuming grouping algorithm [90]. However, it is still unclear which key-points are beneficial for bounding box regression. To address this, an in-depth analysis of the role of each type of key-point is necessary. This could involve a comprehensive evaluation of the importance of each key-point in bounding box regression or the development of a new algorithm that automatically selects the most beneficial key-points.

3.3 Recent techniques in proposal generation

Object proposals are the main results of the first stage in two-stage object detection pipelines. Based on object representations, proposal generation network takes corresponding features generated by the feature extraction network as inputs to generate proposals. Proposals are used by the second stage to generate final predictions. An ideal proposal generation method should generate as few proposals as possible while covering all object instances in the input image. To generate proposals, early object detection methods are usually based on merging super-pixels (e.g., Selective Search [46]) and sliding windows (e.g., Edge Boxes [47]). Although these approaches have been broadly used as the proposal generation methods of choice by many object detectors, they exhibit issues and limitations. First, they require significant computation to process millions of proposals per image. Second, since these methods have no learnable parameters, they are external modules independent of the detector. To solve these issues, Ren et al. [30] introduced novel region proposal network (RPN) which has become the paradigm for designing two-stage object detection pipeline. Figure 9 (a) illustrates the structure of RPN. It includes a 3×3 convolution layer followed by two parallel 1×1 convolution layers for regression and classification. For each anchor $B=\left\{x_B, y_B, w_B, h_B\right\}$, where $\left\{x_B, y_B\right\}$ is the center point of the anchor and $\left\{w_B, h_B\right\}$ is the width and height of the anchor, the classification branch outputs two predictions: the score of it being background and the score of it being foreground. The regression branch aims to predict the transformation ∆ from the anchor B to the target ground truth bounding box $G=\left\{x_G, y_G, w_G, h_G\right\}$ represented as follows:

$\Delta_x=\left(x_G-x_B\right) / w_B, \Delta_y=\left(y_G-y_B\right) / h_B$             (1)

$\Delta_w=\log \left(w_G / w_B\right), \Delta_h=\log \left(h_G / h_B\right)$             (2)

RPN has been used as a proposal generation network in many deep learning-based object detectors. However, RPN is weak at locating objects with extreme shapes or objects in difficult environments because of information loss caused by pooling layers in CNN structure and object representations with predefined shapes. To tackle the shortcomings of RPN, Zhong et al. [48] proposed a cascade architecture in proposal generation stage to improve score and location of proposals (Figure 9 (b), denoted as Iterative RPN in this paper). The proposed cascade structure includes two RPNs. The second RPN takes proposals produced by the first RPN as inputs and further classifies and regresses to generate high-quality proposals. The cascade structure improves the localization capability for objects with various sizes. However, since the anchor location and shape change after each RPN, this structure causes mismatch between anchor boxes and their representations. Another approach, Cascade RPN [15] (Figure 9 (c)) employs one anchor box at each spatial location on a feature map and performs box refinement through multi-stage refinement scheme. In the first refinement stage, dilated convolution layer is used to produce anchor with more semantic information. Since only one anchor is defined at each location in the first stage, Cascade RPN uses the center of the anchor and ground truth box to define positive anchor instead of IoU threshold. Based on the regressed box and features produced by the first refinement stage, the second refinement stage with adaptive convolution layer is used to classify and regress each active bounding box to produce final proposals. To improve the localization capability of RPN, Qiu et al. [49] introduced BorderRPN (Figure 9 (e)) which adds two border alignment modules (BAM) into the original RPN. The first BAM produces border offsets which are then combined with coarse box offsets generated by RPN to enhance bounding box locations. The second BAM produces border classification scores which replace bounding box scores generated by RPN. In both BAM branches, BorderAlign [49] module is adopted to extract and exploit border features which are crucial for achieving better detection accuracy.

Figure 9. The structure of proposal generation networks: (a) RPN; (b) Iterative RPN; (c) Cascade RPN; (d) BorderRPN; (e) CPN; (f) GA-RPN; (g) SC-RPN