kujason / avod

Code for 3D object detection for autonomous driving
MIT License
939 stars 347 forks source link

variation of the number of clusters doesn't work #177

Closed patelrahil224 closed 4 years ago

patelrahil224 commented 4 years ago

Hi,

I tried changing the number of cluster for Pedestrian, Cyclist and People (Number of clusters 3 and 5). Every time after changing the number of cluster new label_cluster is generated with expected values. Training can also performed without any issue. But it seems the network can not permorm any detection as the values in the tensor flow for almost always remains zero.

image

Training and AVOD loss goes down significantly but during the evaluation and inference the network can not detect anything.

image

I am facing trouble figuring out what could be the cause, any suggestions will be helpful

Thanks, Rahil