76.4 map on voc 2017, 22.4 map on coco. 23.6 FPS on iphone 8, 125 FPS on TX2.
一、 key features of pelee:
1. two-way dense layer
2. stem block (improve the feature expression ability without adding computational cost too much)
3. dynamic number of channels in bottleneck layer
DenseNet bottleneck in the module, the number of channels is to increase 4 times, herein, the number of channels is resized according to a dynamic input.
4. transition layer without compression
DenseNet the compression factor in reducing the expression characteristics, herein, transition layer output channels and input channels are left unchanged
5. composite function
Using post-activation (conv-bn-relu) structure, so that the merge CONV bn may be, the use of shallow and wide structure to compensate for the effects caused by this change.
二、optimize SSD and combine SSD with peleeNet
1. feature map selection
Use 19x19, 10x10, 5x5, 3x3, 1x1 size feature map, 38x38 SSD does not use the
2. residual prediction block
Prior to predict each thirds plus a residual block
3. architecture
