search

Epiphqny / Yolact_fcos

YOLACT: Real-time Instance Segmentation on the FCOS detector (without bbox cropping), achives 35.2mAP on coco val
90 stars 19 forks source link

about proto_tower #2

Open Hanawh opened 4 years ago

Hanawh commented 4 years ago

proto_tower is stacked by 32 convolutions according to your code (fcos/modeling/fcos/fcos.py Line 166). Did you try less number convolutions? I have also recently combined yolact and fcos, but my accuracy is only 10

lxtGH commented 4 years ago

@Hanawh Did you try less convolution on prototype generation ? Aslo Hi @Epiphqny Why you use 32 convs to generate prototype ? This make the system run very very slowly also with huge memory cost.

Epiphqny commented 4 years ago

@lxtGH I have not tried less number convolutions, but I think the convolution number in the middle layer is not very important, so you could try less number.