Object Detection

Object Detection

• Object classification
• What is the object in an image
• Classifier → output: class_score[K, N]
• K: The number of predicted objects
• N: The number of class

• Object localization
• What and where is the “single” object in an image
• Regressor → output: [K+4]
• K: The number of predicted objects
• N: The number of class

• Object detection
• What and where is the “multiple” object in an image
• How to detect multiple object? Sliding Window method
• Multiple object localization
• Computation cost is very high. (Inefficient)
• Overfeat paper
• Classifier → output: class_score[K, N]
• Regressor → output: bounding box offset[K, 4]

• Network architecture of Classification and Object Detection
• Final layer shape is different
• Classification: [1, 1, class_num]
• Object detection: [H, W, class_num + box_offset + confidence] (one stage)

• One stage detection & Two stage detection
• One-stage dection
• SSD, YOLO, …
• Backbone
• Feature extractor
• Extract features from input image
• The deeper the layer, the more abstract the feature maps
• Neck
• Merge the different resolution feature maps
• Concatenate/add different scale feature maps
• Dense Prediction
• Predict score of object and bounding box
• Regression layers
• Two-stage dection
• Faster-RCNN, …
• More accurate than one-stage, but high computation cost
• 1st forward
• Get the object candidate regions
• 2nd forward
• Classify the object in region proposals
 

How to find the object

• Grid
• the header layer’s final feature map
• If feature map size is [13, 13], the grid size is [13, 13]
• Predict objects in each grid cell

• Anchor
• The detector which is predict single bounding box
• Predict one object per anchor
• Pre-defined bounding box shape
• In a grid cell, there are serveral anchors

• Bouding box

• objectness score
• Object or not

• Class score
• cat or dog or car …

Loss function

• Softmax

• Cross-entropy loss

• MSE loss, MAE loss

Terms

• IOU
• Interection over union
• The mertic of how well predicts the bounding box compared with GT box
• ex) IOU > 0.5, positive box. Otherwise, negative box

• NMS
• Non-Maximum Suppression
• Filtering the best predicted boxes using IOU and confidense score

• Data annotaion
• Draw bounding box on the object and labeling its class
• Object detection dataset
• One image, one GT
• GT : Ground Truth, annotating the object information(bounding box, class etc)
• Training & Evaluation
• Training set: using when training the model
• Evaluation set: using when evaluating the trained model
• Why evaluate the trained model?
• Training the model using training set, the model fitted on the training set domain
• Evaluation set generalizes the model, prevents “overfitting” the model
• Tools
• labelImg
• Yolo_Label

One-stage object detectors

• YOLO
• You Look Only Once
• Joseph Redmon (2015)
• YOLOv1, YOLOv2, v3, …, v5, YOLOR
• Backbone
• GoogleLenet(YOLOv1), Darknet19, Darknet53

• SSD
• Single Shot Detectors
• Wei Liu (2015)
• Backbone
• VGG-16

Two-stage object detectors

• RCNN
• Selective Search, Rule-based region proposals
• 200 region proposals
• Proposals resize 227*227
• Bouding box: regression
• Class: SVM classifier

• Fast - RCNN
• ROI pooling
• Not SVM classifier, Softmax regression
• Bouding box: regression
• Not Selective search, Region proposal Network

Evaluation metric

• Precision (정밀도)
• the ratio of true positive (true predictions) (TP) and the total number of predicted positives (total predictions)

• Recall (재현율)
• the ratio of true positive (true predictions) and the total of ground truth positives (total number of cars)

• Terms
• TP(True Posivite) : True prediction
• FP(False Positive) : False prediction
• FN(False Negative) : not predict GT
• TN(True Negative) : not predict non-GT

• F1 score
• precision과 recall의 조화평균
• high recall + high precision
• the class is perfectly handled by the model
• low recall + high precision
• the model can’t detect the class well but is highly trustable when it does
• high recall + low precision
• the class is well detected but the model also include points of other classes in it
• low recall + low precision
• the class if poorly handled by the model

• PR curve
• Depending on the object confidence threshold, draw the PR curve

• mAP
• Mean average precision, PR curve area

• Confusion matrix
• Mesuarement of True positive and False positive between classes

YOLOv3

• Architecture

• Configure file
• hyperparameters
• Model architecture
• Data augmentation
• Input resolution

• Annotaion labeling format
• YOLO labeling format
• Class_index x_center y_center width height
• x y w h is normalized value [0, 1]