A Survey of Deep Learning-based Object Detection

2021/12/15

the purpose of object detection: locating instances of semantic objects of a certain class

*object detection and domain-specific object detection

most of the state-of-the-art object detectors utilize deep learning networks as their backbone and detection network to extract features from input images (or videos), classification and localization respectively

well-researched domains of object detection include multi-categories detection, edge detection, salient object detection, pose detection, scene text detection, face detection and pedestrain detection etc

*benchmark: 一个领域公认的基准，具体表现为领域中论文一致使用的数据集、评价指标

• two kinds of object detectors

  two-stage: Faster R-CNN; one-stage: YOLO

  two-stage detectors have high localization and object recognition accuracy, whereas the one-stage detectors achieve high inference speed

most of backbone networks for detection are the network for classification task taking out the last FC layer

2021/12/16

Two-stage Detectors

• R-CNN (first deep learning-based detector)
• Fast R-CNN (use of RoI Pooling)
• Faster R-CNN (use of region proposal network/RPN, the use of multi-scale anchors)
• Mask R-CNN (for instance segmentation task, use of feature pyramid network/FPN, use of RoIAlign)

*N+1-way classification layer, N for object classes and 1 for background

One-stage Detectors

• YOLO (real-time detection of full images and webcam)
• YOLOv2 (adopt a series of design decisions from past works with novel concepts， new backbone)
• YOLOv3 (an improved version of YOLOv2)
• SSD (a single-shot detector for multiple categories)
• DSSD (a modified version of SSD)
• RetinaNet (use of focal loss)
• M2Det (have no idea about this)
• RefineDet (have no idea about this)

detecting an object has to state that an object belongs to a specified class and locate it in the image

the localization of an object is typically represented by a bounding box

benchmarks

• PASCAL VOC dataset (basic)
• MS COCO benchmark (large in images per class)
• ImageNet (large in class num)
• VisDrone2018 (have no idea about this)
• OpenImages V5 (have no idea about this)
• Recall
• Precision
• Average Precision (AP)
• mean Average Precision (mAP)

deep neural network based object detection piplines:

• image pre-processing: resize raw data and perform data augmentation
• feature extraction: a key step for further detection
• classification and localization: concluding classification scores and bounding box coordinates
• post-processing: delete any weak detecting results (like NMS)

to obtain precise detection results, there exists several methods can be used alone or in combination with other methods:

• Enhanced features: for extracting effective features from input images (like FPN, Attention)
• Increasing localization accuracy: design a novel loss function
• Solving negatives-positives imbalance issue: for one-stage, like hard mining / add some item in classification loss
• Improving post-processing NMS methods
• Combining one-stage and two-stage detectors to make good results
• Complicated scene solutions (have no idea about this)
• Anchor-free: still a novel direction for further research
• Training from scratch: 有的数据集就是需要从头训练才能保证稳定以及准确性
• Designing new architecture
• Speeding up detection
• Achieving Fast and Accurate Detections

typical application areas:

• Security Field: Face detection, Pedestrain detection, Anomaly detection
• Military field: Remote sensing OD
• Transportation field
• Medical field: Computer Aided Diagnosis (CAD) systems
• Life field: Pattern detection, Image caption generation