PaperlayerView exact revision on arXiv

Source-linked AI summary

A Review of Object Detection Models based on Convolutional Neural Network

F. Sultana, A. Sufian, P. Dutta

arXiv:1905.01614v3cs.CV

TL;DR

Object detection requires estimating both object classes and locations, and CNN-based models have become a major approach. This chapter reviews CNN detectors by two-stage and one-stage designs, describing their architectures and training details and comparing their performance. It concludes that two-stage models are more accurate but slower, while successive architectural and loss-function improvements support more accurate and faster real-time detection.

  • Problem

    Object detection requires estimating the class and location of every object instance in an image.

  • Method

    The chapter reviews CNN-based object detectors by two-stage and one-stage approaches, describing architectures and training details and comparing performance across datasets.

  • Results

    Two-stage models achieved higher object-detection accuracy than one-stage models but were slower; RefineDet combined merits of both approaches and achieved state-of-the-art performance.

  • Takeaways & Limitations

    Improvements in CNNs, detection architectures, pooling methods, and loss functions support progress toward more accurate and faster real-time object detection.

  • Takeaways & Limitations

    R-CNN compromises image scale, aspect ratio, or originality because CNN inputs must be fixed-sized while its region proposals are arbitrary.

Abstract

from arXiv · show

Convolutional Neural Network (CNN) has become the state-of-the-art for object detection in image task. In this chapter, we have explained different state-of-the-art CNN based object detection models. We have made this review with categorization those detection models according to two different approaches: two-stage approach and one-stage approach. Through this chapter, it has shown advancements in object detection models from R-CNN to latest RefineDet. It has also discussed the model description and training details of each model. Here, we have also drawn a comparison among those models.

1 Introduction

Object detection estimates both the classes and locations of every object instance in an image. This review examines CNN-based detection models and their architectures and training details.

  • Object detection estimates object classes and locations for every instance in an image, unlike image classification.
  • Before deep learning, handcrafted features such as SHIFT and HOG supported models that reached around 35% mAP on PASCAL VOC.
  • The review covers different state-of-the-art CNN-based object detection models.
  • It describes model architectures and training details, compares performance across datasets, and concludes the chapter.

2 Different Object Detection Models

CNN-based object detectors are organized into two-stage and one-stage approaches. The two-stage approach first generates object proposals and then classifies and localizes them.

  • CNN-based object detection models are categorized into two-stage and one-stage approaches.
  • Two Stage approach: Two-stage detectors generate region or object proposals first, then classify those proposals and detect them using bounding boxes.
  • R-CNN: R-CNN was the first CNN-based two-stage detector, using selective search to generate approximately 2000 class-independent proposals per image.
  • R-CNN: R-CNN used an AlexNet-like CNN to extract fixed-length features from warped region proposals, followed by classification and detection processing.
  • R-CNN: R-CNN training used ILSVRC-2012 pretraining, a 21-way classifier, SGD, and an IoU threshold of ≥0.5 for positive proposals.
  • R-CNN: R-CNN’s fixed-size CNN inputs required cropping, warping, or predefined scales, compromising region scale, aspect ratio, or image originality.

SPP-net:

SPP-net inserts spatial pyramid pooling between convolutional and fully connected layers to process variable-sized region proposals. This increases scale invariance, reduces overfitting, and makes detection faster than R-CNN.

  • SPP-net inserts a spatial pyramid pooling layer between convolutional and fully connected layers.
  • SPP layers convert variable-length region-proposal features into fixed-length outputs for fully connected layers.
  • SPP-net supports training and testing with varying image sizes and scales, increasing scale invariance and reducing overfitting.
  • SPP-net applies the CNN once to the entire image and extracts region proposals from the final convolutional feature map.
  • SPP-net is much faster than R-CNN because it reuses the shared convolutional feature map for region proposals.

Fast R-CNN:

Fast R-CNN replaces R-CNN’s multistage, expensive pipeline with single-stage joint learning for classifying region proposals and refining their locations. It processes the entire image once and reports substantial speed gains over R-CNN and SPP-net.

  • Fast R-CNN: Fast R-CNN jointly classifies region proposals and corrects their spatial locations through a single-stage training algorithm.It was designed to address multistage training, expensive computation, and slow detection in R-CNN and SPP-net.
  • Fast R-CNN: 9× faster than R-CNN and 3× faster than SPP-net, Fast R-CNN uses one CNN pass over the entire image before processing RoIs.RoI pooling reshapes selected regions into fixed-length feature vectors for subsequent fully connected layers.
  • Fast R-CNN: Fast R-CNN was tested with ImageNet-pretrained CaffeNet, AlexNet, VGG CNN M 1024, and VGG16 networks.The networks were modified by replacing the final pooling, fully connected, and softmax components with Fast R-CNN components.
  • Fast R-CNN: RoI pooling and training techniques helped Fast R-CNN achieve higher accuracy while reducing time complexity relative to earlier models.The model uses a CNN feature map, RoI pooling, and two sibling output branches.

Faster R-CNN:

Faster R-CNN removes the separate region-proposal bottleneck by replacing proposal algorithms with a trainable region proposal network. The section also introduces FPN’s multilevel features for scale handling and connects this progression to Mask R-CNN.

  • Faster R-CNN: Faster R-CNN replaces time-consuming region proposal algorithms with a fully convolutional region proposal network that outputs candidate boxes and objectness scores.The RPN accepts images of arbitrary size and determines whether proposals contain objects.
  • Faster R-CNN: The RPN predicts proposals from convolutional feature maps using sliding-window anchors and an objectness threshold.Selected anchors and feature maps are then passed to the later detection stages.
  • FPN: FPN addresses scale and localization issues by combining bottom-up backbone features with a top-down multilevel feature pyramid.Its region proposals and RoIs use features from different pyramid levels to combine multi-scale information.
  • FPN: FPN uses a ResNet-based Fast R-CNN backbone and is trained end-to-end on MS COCO with 80 categories.Training used synchronized SGD with two images and 256 anchors per image.
  • Mask R-CNN: Mask R-CNN extends Faster R-CNN toward instance segmentation by using RoIAlign to align extracted features with object locations.It retains Faster R-CNN’s region proposal network while predicting object-instance masks rather than only bounding boxes.

Mask R-CNN:

Mask R-CNN extends Faster R-CNN to instance segmentation by retaining its proposal network and replacing RoI pooling with RoIAlign. Its implementation uses ResNet or ResNeXt with FPN as the backbone.

  • Mask R-CNN: ResNet or ResNeXt combined with FPN forms the base architecture, with RoIs extracted from different FPN levels at varying scales.The authors report that ResNet and FPN improved accuracy and speed.
  • Mask R-CNN: Training used a 1:3 positive-to-negative RoI ratio, a learning rate of 0.02 for 160K iterations, and a tenfold reduction for the next 120K iterations.Momentum and weight decay were 0.9 and 0.0001, respectively.

2.2 One stage approach:

One-stage detectors perform classification and regression in a single shot using dense sampling across locations, scales, and aspect ratios. The reviewed models include YOLO, SSD, YOLO9000, and RefineDet.

  • One stage approach: One-stage detectors perform classification and regression in a single shot using regular and dense sampling over locations, scales, and aspect ratios.This contrasts with two-stage detection, where proposals are generated before classification and regression.
  • YOLO: YOLO directly predicts class probabilities and bounding boxes from the input image using a simple CNN.The image is divided into a fixed grid, whose cells predict boxes with confidence scores based on object probability and IoU.
  • YOLO: YOLO training used 24 convolutional layers, two fully connected layers, batch normalization, leaky rectified linear activations, dropout, and heavy data augmentation.The final layer used linear activation, while the other layers used leaky rectified linear activation.
  • SSD: SSD processes the entire image through multiple convolutional layers and predicts boxes from feature maps at different network positions.Its anchor boxes encode center coordinates, width, and height, while extra feature layers use 3 × 3 filters.
  • SSD: SSD matches anchors to ground-truth boxes using best Jaccard overlap and retains anchors with overlap greater than 0.5.Feature maps from different layers help handle objects at different scales.
  • YOLO9000: YOLO9000 is described as a real-time object detector capable of detecting more than 9000 object categories.

YOLO9000:

YOLO9000 improves the original YOLO system through joint training across ImageNet and MS COCO while preserving speed. RetinaNet instead uses a one-stage FPN detector with classification and box-regression subnetworks, trained using focal loss for class imbalance.

  • YOLO9000: YOLO9000 jointly trains on more than 9000 classes from ImageNet and MS COCO detection data.The method combines data sources using a WordTree hierarchy.
  • YOLO9000: YOLO9000 uses batch normalization, 416 × 416 inputs, removes a pooling layer, and uses anchor boxes for bounding-box prediction.
  • RetinaNet: RetinaNet is a unified one-stage detector that densely samples object locations using an FPN backbone.Its two task-specific sub-networks perform classification and bounding-box regression at each FPN level.
  • RetinaNet: RetinaNet uses focal loss to address extreme foreground–background class imbalance during one-stage detection training.The total training loss combines focal loss with smooth L1 loss for box regression.

RefineDet:

RefineDet is a single-shot convolutional detector organized around anchor refinement and object detection modules. The ARM filters negative anchors and adjusts them before the ODM predicts refined locations, sizes, and class labels.

  • RefineDet: RefineDet is a single-shot detector based on a feed-forward convolutional network with two interconnected modules.The modules are the anchor refinement module and the object detection module.
  • RefineDet: The ARM filters negative anchors and roughly adjusts anchor sizes and locations to initialize the successive regressors.
  • RefineDet: The ODM uses refined anchors to improve regression accuracy and predict multiple class labels.
  • RefineDet: RefineDet uses pretrained VGG-16 and ResNet-101 base networks, with added convolutional or residual layers for higher-level and multiscale information.

3 Comparative Results

The review compares object detectors using accuracy and speed, but emphasizes that differences depend on architecture, training, matching, suppression, and other design choices. Table 1 reports mAP across PASCAL VOC and MS COCO and summarizes speed and organizational approach.

  • 3 Comparative Results: Performance comparison is difficult because the reviewed models use two-stage and one-stage approaches with differing design choices.Relevant factors include feature extractors, input resolution, matching strategy, IoU threshold, suppression, proposals, encoding, losses, optimization, and hyperparameters.
  • 3 Comparative Results: Table 1 reports object-detection accuracy using mean average precision across PASCAL VOC and MS COCO datasets.
  • 3 Comparative Results: The table’s final columns provide an overview of model speed and organizational approach.

4 Conclusion

The chapter concludes that CNN-based object detectors have progressed through two-stage and one-stage approaches, with a recurring trade-off between accuracy and speed. It attributes progress to stronger CNNs, new architectures, pooling methods, and loss functions, culminating in RefineDet’s reported state-of-the-art performance.

  • 4 Conclusion: Two-stage models achieved higher object-detection accuracy than one-stage models but were slower.
  • 4 Conclusion: RefineDet combined merits of two-stage and one-stage approaches and achieved state-of-the-art performance.
  • 4 Conclusion: The review links model progress to better CNNs, new detection architectures, pooling methods, and novel loss functions.
  • 4 Conclusion: The chapter presents improvements in object detectors as supporting the pursuit of more accurate and faster real-time object detection.
Loading 1905.01614v3…