when inferring the real shot picture, a target has multiple detected bounding box of different categories.

zylo117 / Yet-Another-EfficientDet-Pytorch

The pytorch re-implement of the official efficientdet with SOTA performance in real time and pretrained weights.

GNU Lesser General Public License v3.0

5.21k stars 1.27k forks source link

when inferring the real shot picture, a target has multiple detected bounding box of different categories. #622

Open CaiyuZhang opened 3 years ago

CaiyuZhang commented 3 years ago

Inferring the pictures in the dataset, the effect is not bad. But when inferring the real shot picture, a target has multiple detected bounding box of different categories. Below is the training results. It is not that bad, right? I noticed that the indicators for the small and medium goals are both -1. So I increased the proportion of the target in the picture, but it got worse result, nothing has been detected.

zylo117 commented 3 years ago

Set higher confidence threshold and nms threshold

CaiyuZhang commented 3 years ago

Set higher confidence threshold and nms threshold

Thanks to reply! Appreciate your great work. Setting a higher confidence threshold will result in no result output, and nms threshold is to filter those bounding box belong to same class, my case is one target has multiple box belong to different classes.

zylo117 commented 3 years ago

Then you have to keep training the model until it performs better