About the roi align layer

[suzhenghang]

Hi @jasjeetIM ,
After adding the roi align layer，I replace the roi pooling layer in the trainval.prototxt like the following:

...................
top: "conv5_3"
}
layer {
name: "roi_align5"
type: "ROIAlign"
bottom: "conv5_3"
bottom: "rois"
top: "pool5"
roi_align_param {
pooled_w: 7
pooled_h: 7
spatial_scale: 0.0625 # 1/16
}
}
layer {
name: "fc6"
...................

Did it right? If so, did you compare the roi align layer and roi pooling layer in object detection? Could it increase the mAP?

Thanks in advance.

[alphacentauri12]

Hi @suzhenghang / @jasjeetIM
have you made a comparison "RoIAlign vs. RoIPool" layer?
I want to replace the RoIPool layer in Faster R-CNN with the RoiAlign layer, hoping to get a more precise detection. Do you think that it is possible?

[ghost]

Hi@suzhenghang , could you let us know the update please? thank you