The belndmask is not so good

[NingNanXin]

Hi, I test the blendmask and MaskRcnn in detectron2. The first image is the original image, the second is MaskRcnn, the third is belndmask. And you can see the belndmask is not so good. I guess the multi-scale problem is not effectively solved. And the attention is not very useful. Maybe we can adjust it?
If anyone has a better solution, welcome to discuss it with me. Thanks a lot!

[cty-xiaobai]

I also encountered the same problem. The scores for detecting people or cars are very low

[NingNanXin]

I also encountered the same problem. The scores for detecting people or cars are very low

Hi, cty-xiaobai!
In the past 4 days, I Retrained the blendmask and MaskRcn on COCO, and the result is the same. But the blendmask is based on detecron2, the detect module should be the same as MaskRcnn. However, I am busy with my work and don't have enough time to find the causes. Could you let me know if you find the causes? Also, I will research the network when I am free. Thanks a lot!

[cty-xiaobai]

I also encountered the same problem. The scores for detecting people or cars are very low

Hi, cty-xiaobai!
In the past 4 days, I Retrained the blendmask and MaskRcn on COCO, and the result is the same. But the blendmask is based on detecron2, the detect module should be the same as MaskRcnn. However, I am busy with my work and don't have enough time to find the causes. Could you let me know if you find the causes? Also, I will research the network when I am free. Thanks a lot!

Hi,have you ran the evaluation program? The value of AP should be correct

[Tiberius131]

ýou need to set the threshold value lower. E.g --confidence-threshold 0.20

[kaixin-bai]

I followed the README of BlendMask, but cannot get the right result.

[pzh11001]

I also found this problem, especially on small data sets. I built my own data set and used coco pre training model. There are only 200 training pictures. Using maskrcnn to train 50 iters can quickly achieve high recognition rate. But with blendmask, even if you train more than 200 iters, you can only get a small score. And some objects in the scene can't be recognized at all.

[Joinjiang]

i have the same problems, and the blendmask is in poor performance on my single object dataset, how to improve its performance, have any idea?