search

traveller59 / second.pytorch

SECOND for KITTI/NuScenes object detection
MIT License
1.72k stars 721 forks source link

Separating RPN from VoxelNet #295

Open themousepotato opened 4 years ago

themousepotato commented 4 years ago

Hi,

I'm trying to separate RPN from the VoxelNet and make it as a different model. But, on trying to do so, while loading an embedding which is generated by MiddleFeatureExtractor to the RPN, it's not outputting the correct results. Any idea how can I make it work?

Thanks in advance.

themousepotato commented 4 years ago

Closed by mistake. Can someone please help me with this?

darshb34 commented 4 years ago

128 feature maps from the MFE are already being fed to RPN to predict the 3d and bev bbox.

  1. what do you mean by not outputting correct results? I am facing similar problem, the predictions are empty tensors.
themousepotato commented 4 years ago

I'm trying to achieve another goal. I need the part till middle feature extractor(including that) and the part following that as two separate models so that I can train them differently. I'm not getting accurate bounding boxes when I split the current model into two and load pretrained weights. I'm not getting empty tensors, but values which are not accurate.