Object Detection Develop History-3
==One Stage Method ==
stage 1 = region proposal + feature extract + bounding box
YOLO
SSD
YOLO V1
有B種物體
輸入影像透過CNN得到FEATURE MAP
將照片分成SxS的格子
若物體的中心掉在格子內，則由該格去算對每個物體的信心分數(B個)
並且計算物體座標
Image Not Showing
Possible Reasons
- The image file may be corrupted
- The server hosting the image is unavailable
- The image path is incorrect
- The image format is not supported
Learn More →
YOLO 誤差
與Fast R-CNN相比，YOLO對背景的辨識率很高
但是物體的定位誤差比較大
定位誤差
相鄰物體過多，yolo較難偵測
因為每個格子只能輸出兩個物體
SSD
Single Shot MultiBox Detection
一樣是one stage的做法，但是改良了yolo的一些缺點
yolo 對小物體偵測不佳，SSD則加入Multi-scale
使用預先設定好的bounding box進行預測，加快訓練速度
且IOU設定為0.5以上才算是有偵測到
Image Not Showing
Possible Reasons
- The image file may be corrupted
- The server hosting the image is unavailable
- The image path is incorrect
- The image format is not supported
Learn More →
==物體大小與FEATURE MAP 關係 ==
比較大的物體 –-> 在低解析度,ex. 4x4 大小的feature map 可以偵測
比較小的物體 –-> 在高解析度,ex. 8x8 大小的feature map 可以偵測
Image Not Showing
Possible Reasons
- The image file may be corrupted
- The server hosting the image is unavailable
- The image path is incorrect
- The image format is not supported
Learn More →
因此SSD拿掉YOLO最後的Fully connect層，改用multi-scale network
來解決不同大小物體的分類跟定位問題
Image Not Showing
Possible Reasons
- The image file may be corrupted
- The server hosting the image is unavailable
- The image path is incorrect
- The image format is not supported
Learn More →
更多的YOLO vs SSD
HackMD
