Precision is always very low when training

[alex-razor]

i train the model on my custom gray scale data.

However, i always get very low precision (in the following its the highest i could get):

EVAL: Class 0: Recall: 0.9116, Precision: 0.0650, AP: 0.7955
EVAL: Recall: 0.9116, Precison: 0.0650, mAP: 0.7955
EVAL: loss: total: 3.58, xy: 0.29, wh: 0.29, conf: 2.99, class: 0.01

Is it because -

• something's wrong with my data?
• i don't train properly?
• some bug?

[alex-razor]

Ok, so i figured out that it's because i have at most 2 objects in a frame and nms_topk=150 # keep at most nms_topk outputs after nms.

[JonasHendl]

Hey, what did you change to fix this? I might have the same problem. Thanks

[yjsdut]

Hello, I have encountered the same problem. Have you solved it?How to solve it?Thanks!

[Byronnar]

Hello, I have encountered the same problem. Have you solved it?How to solve it?Thanks! 我解决了。如果你不是训练的coco，你只需要把nms_topk调低就好了。因为自己的训练集类别数没有80

[Byronnar]

Hey, what did you change to fix this? I might have the same problem. Thanks

我解决了。如果你不是训练的coco，你只需要把nms_topk调低就好了。因为自己的训练集类别数没有80

[alex-razor]

I am not sure if it's how it suppose to be, but - i changed nms_topk to 2..now precision is at most 0.5..

[alex-razor]

Hey, what did you change to fix this? I might have the same problem. Thanks

我解决了。如果你不是训练的coco，你只需要把nms_topk调低就好了。因为自己的训练集类别数没有80

I have only 1 class, data is like COCO..what are you suggesting to do?

[duducode]

Ok, so i figured out that it's because i have at most 2 objects in a frame and nms_topk=150 # keep at most nms_topk outputs after nms.

你好该怎么修改呢，比如我的类别是10类

[tachikoma777]

Is nms_topk only have effect on eval or also on trainning? Should I retrain the network with a new nms_topk?

[iodncookie]

i might have the same problem .please save kid.

[mahaweilu]

I am not sure if it's how it suppose to be, but - i changed nms_topk to 2..now precision is at most 0.5..

The same as mine. My class number is 1. I changed the nms_topk to 10 and the most is 0.4. Did you improve the precision result by changing other parameters?