Faster R-CNN: Model Design¶
Architecture¶
Faster R-CNN Model Structure:
+-------------------------------+
| Input Image |
+-------------------------------+
|
|
|
|
|
+-------------------------------+
| Convolutional Backbone |------------------|
+-------------------------------+ |
| |
| |
| |
| |
| |
+-------------------------------+ |
| Region Proposal Network (RPN) | |
+-------------------------------+ |
| |
| |
|----------------------------------|
|
|
+-------------------------------+
| RoI Pooling Layer |
+-------------------------------+
|
|
|
|
|
+-------------------------------+
| Fully Connected Layers |------------------|
+-------------------------------+ |
| |
| |
| |
| |
| |
+-------------------------------+ +-------------------------------+
| Classification | | Localization |
+-------------------------------+ +-------------------------------+
where:
both the RoI pooling layer and its following-up FC layers constitute the Fast R-CNN detector [4]
the backbone convolutional network adapts
ZF
orVGG
[16]
Back to Object Detection.