In computer vision, object detection is one of the tasks that AI shows remarkable result. Each object detection task has different models and creates different level of output. This is my wrap-up of object detection tasks.
Image Classification
Image Classification detects class from overall image input and returns class with probability.
classes = [“dog”, “cat”, “nothing”]
prediction = [ 0.8 , 0.1 , 0.15]
Models
Xception
VGG
ResNet
MobileNet
DenseNet
Model Validation
Confusion Matrix, Accuracy, Precision, Recall, F-measure
Image Classification・Localization
Image Classification with Localization creates windows to detect class from image input and returns class, probability and bounding box.
# class label
classes = [“dog”]
prediction = [ 0.8]
# Bounding Box
legend = [ “X-Position”, “Y-Position”, “Length”, Height”]
prediction = [130, 285, 100, 185]
Models
HOG
BOF
PASCAL VOC
Model Validation
mAP(mean Average Precision)、IoU(Intersection over Union)
Object Detection
Object Detection creates windows to detect multiple objects and returns those classes, probability and bounding box.
# class label
classes = [“dog”, “dog”]
prediction = [ 0.98, 0.97]
# Bounding Box
legend = [ “X-Position”, “Y-Position”, “Length”, Height”]
prediction = [130, 285, 100, 185], [130, 285, 100, 185]
Models
R-CNN
Fast R-CNN
SSD: Single Shot MultiBox Detector
Mask R-CNN
YOLO
Model Validation
mAP(mean Average Precision)、IoU(Intersection over Union)
Semantic Segmentation
Semantic Segmentation detects and provides class at pixel level. Same class is detected as the same object, so we cannot count the target of a number.
Models
U-NET
MULTISCALE
HYBRID CNN-CRF
Model Validation
IoU and per-pixel accuracy
Instant Segmentation
Instant Segmentation detects objects in the image and assigns a class at pixel level. Compared to the semantic segmentation, we can count the target of a number.
Models
Mask R-CNN
DeepMask
FCIS
Model Validation
Average precision over different IoU thresholds
Panoptic Segmentation
Panoptic Segmentation detects objects at pixel level and assigns class of objects and background.
Models
Panoptic Feature Pyramid Network
UPSNet
Model Validation
Panoptic Quality (PQ)
Summary
担当者:HM
香川県高松市出身 データ分析にて、博士(理学)を取得後、自動車メーカー会社にてデータ分析に関わる。その後コンサルティングファームでデータ分析プロジェクトを歴任後独立 気が付けばデータ分析プロジェクトだけで50以上担当
理化学研究所にて研究員を拝命中 応用数理学会所属