1.The original size of pool_h and pool_w is 6.6. like:
layer {
name: "roi_pool_conv5"
type: "ROIPooling"
bottom: "conv5"
bottom: "rois"
top: "roi_pool_conv5"
roi_pooling_param {
pooled_w: 6pooled_h: 6
spatial_scale: 0.0625 # 1/16
}
}
2.Then I change the size，like:
layer {
name: "roi_pool_conv5"
type: "ROIPooling"
bottom: "conv5"
bottom: "rois"
top: "roi_pool_conv5"
roi_pooling_param {
pooled_w: 8pooled_h: 8
spatial_scale: 0.0625 # 1/16
}
}
And get the bad results. The map is reduced from 0.8 to 0.117. Any one know?
Hi @315386775 : I am curious to know if you were able to find out the root cause of low map? I am also experimenting with irregular shaped output from ROIPooling layer based on changes done in R2CNN. and my final detections are totally screwed up.
ravikantb
commented
6 years ago
1.The original size of pool_h and pool_w is 6.6. like: layer { name: "roi_pool_conv5" type: "ROIPooling" bottom: "conv5" bottom: "rois" top: "roi_pool_conv5" roi_pooling_param { pooled_w: 6 pooled_h: 6 spatial_scale: 0.0625 # 1/16 } } 2.Then I change the size，like: layer { name: "roi_pool_conv5" type: "ROIPooling" bottom: "conv5" bottom: "rois" top: "roi_pool_conv5" roi_pooling_param { pooled_w: 8 pooled_h: 8 spatial_scale: 0.0625 # 1/16 } } And get the bad results. The map is reduced from 0.8 to 0.117. Any one know?