search

dbolya / yolact

A simple, fully convolutional model for real-time instance segmentation.
MIT License
4.97k stars 1.32k forks source link

Implementation on FCOS #397

Open Epiphqny opened 4 years ago

Epiphqny commented 4 years ago

Hi @dbolya , thanks for your wonderful work, i have implemented the core modules of YOLACT v1 on FCOS in https://github.com/Epiphqny/Yolact_fcos, without bounding box cropping, it can achieve 35.2 mAP on COCO val2017 set with the ResNet-101 backbone.

Auth0rM0rgan commented 4 years ago

Hey @Epiphqny, Nice Job! Would you please tell me what FPS you have achieved? Do you have any plan to implement Yolact++ as well?

Thanks!

Epiphqny commented 4 years ago

Hey @Epiphqny, Nice Job! Would you please tell me what FPS you have achieved? Do you have any plan to implement Yolact++ as well?

Thanks!

The speed is slower than the original yolact, as the base detector(FCOS) is larger, the FPS is about 7.7. As I used serial inference for each instance, there is still room to improve, I will convert it to parallel prediction for each image later.

Auth0rM0rgan commented 4 years ago

I think detectron2 demo is not optimized well at all in terms of the speed. YOLACT visualization tools are optimized very well for better speed. You will achieve higher FPS If you can use yolact visualization tools.

Epiphqny commented 4 years ago

Thanks for your advice, I will try it later.

abhigoku10 commented 4 years ago

@Epiphqny great work in getting FCOS feature any plans to add it on yolact++ since the drop in performance will less compared to yolact

Epiphqny commented 4 years ago

@abhigoku10 I will update the repo if many people request that.