YOLOv1
Yolov1
Paper : https://arxiv.org/abs/1506.02640
Object Detectionģ ėķ ģė”ģ“ ģ ź·¼ė²(one stage object detection) YOLO(You Only Look Once)ź° ģ²ģ ģ ģė ė ¼ė¬øģ ėė¤.
ķµķ©ė 구씰(bounding box + class probability)넼 ź°ģ§źø° ė문ģ ė¹ ė¦ ėė¤.
45 FPS / 155 FPS(Fast)
Unified Detection
S x S grid (S = 7)
B(num of bounding box) = 2
C(num of class) = 20
output tensor : 7 x 7 x (B x 5(x, y, w, h, confidence) + C)
test time

Network Design
GoogLeNet źø°ė° ėŖØėø
Convolution Layer : 24ź°, 9ź°(Fast)
Fully Connected Layer : 2ź°
Loss Function
: Objectź° ģ”“ģ¬ķė Grid Cell
: Predictor Bounding Box
: Objectź° ģ”“ģ¬ķė ź²½ģ° grid cellģ predictor bounding box
: Objectź° ģ”“ģ¬ķģ§ ģė ź²½ģ° grid cellģ predictor bounding box
: Objectź° ģ”“ģ¬ķė ź²½ģ° grid cell
ģ“ėÆøģ§ ėė¶ė¶ģė objectź° ģģ ź²ģ“ź³ confidenceė ģ ė¶ 0ģ¼ė” ģė “ķė ¤ź³ ķ ź² ģ ėė¤. ź·øė” ģøķ“ ė°ģėė gradientź° ė묓 커ģ§ė ķģģ ė§ģ주기 ģķ“ģ ģ¶ź° parameter넼 ģ¬ģ©ķ©ėė¤.
: x, y, w, h lossģ ź· ķģ ģķ parameter. (defalut : 5)
: object lossģ ź· ķģ ģķ parameter. (defalut : 0.5)
x, yģ loss넼 구ķ©ėė¤.
w, hģ loss넼 구ķ©ėė¤. (ź°ė”, ģøė”ģ ģ ź³±ź·¼ģ ģģø”ķ©ėė¤.)
confidence scoreģ loss넼 구ķ©ėė¤. ()
confidence scoreģ loss넼 구ķ©ėė¤. ()
conditional class probabilityģ loss넼 구ķ©ėė¤.
Training
ImageNet 1000-class competition datasetģ¼ė” 20ź°ģ convolution layer, avg pooling layer, fully connected layer넼 ź°ģ§ ėŖØėøģ pretraining ķ©ėė¤. ķ©ėė¤.
randomly initialized weights넼 ź°ģ§ė 4ź°ģ convolution layerģ 2ź°ģ fully connected layer넼 ģ¶ź°ķ©ėė¤.
ģøė¶ģ ģø ģź°ģ 볓넼 ģķ“ ķ“ģė넼 224 x 224ģģ 448 x 448ė” ėė øģµėė¤.
bounding boxģ ķź³¼ ėģ“넼 ģ ź·ķ(0 ~ 1) ķģģµėė¤.
ė§ģ§ė§ Layerģ linear activation functionģ ģ¬ģ©ķģź³ ėėØøģ§ ė¤ė„ø layerģė leaky relu넼 ģ¬ģ©ķ©ėė¤.
parameters
epoch
: 135batch
: 64momentum
: 0.9weight decay
: 0.0005learning rate
: 0.001 -> 0.01 -> 0.001 -> 0.000175 epoch
: 0.0130 epoch
: 0.00130 epoch
: 0.0001
dropout rate
: 0.5data augmentation
random scaling
HSV ģģ ź³µź°ģģ ģµė 1.5ė°° ź¹ģ§ exposureź³¼ saturationģ ģģė” ģ”°ģ ķ©ėė¤.
Inference
one stageė¼ģ ė§¤ģ° ė¹ ė¦ ėė¤.
ģ“미ģ§ė¹ 98ź°ģ bounding boxģ ź° boxģ ėķ class probability넼 ģģø”ķ©ėė¤.
ź° objectė¹ ķėģ bounding boxė” ģģø”ķė¤.
ķ° objectė ģ¬ė¬ź°ģ ģ ģ ķ ėė¦¬ģ ź·¼ģ²ģ ģė 물첓ė ģģø”ķźø° ģ“ė µģµėė¤. NMSė” ķ“ź²°ķ ģ ģģ§ė§ R-CNN ė§ķ¼ ģ±ė„ģ ķ¬ź² ģķ„ģ 미ģ¹ģ§ė ģģµėė¤.
Limitation
Small Objectź° ėŖØģ¬ ģģ¼ė©“ ģ ź²ģ¶ķģ§ ėŖ»ķ©ėė¤.
Localization Errorź° ėģµėė¤.
Benchmark
Yoloė ė¹ ė„“ź³ ź°ė „ķ©ėė¤.
Yolo ģ“ģ ģ ģ¬ģ©ė real time object detection ė³“ė¤ ģ±ė„ģ“ ģ¢ģµėė¤.
Yoloź° Fast-RCNN ė³“ė¤ Localization Errorź° ģ¢ģ§ ģģµėė¤.
Yoloź° Fast-RCNN ė³“ė¤ Background Errorź° ģ¢ģµėė¤.
Yoloģ Fast-RCNNģ ź²°ķ©ķ“ģ ģ¬ģ©ķė©“ ģ¢ģµėė¤.
ķ“ėģ¤ ė³ė” ģ ķė넼 ė¹źµķ ķ ģ ėė¤.
Last updated
Was this helpful?