Closed afonso-sousa closed 3 years ago
Hi @AfonsoSalgadoSousa , As an author of ImVoxelNet can confirm that you are right. This design choice follow the original implementation. However this is probably not the best aggregation method and for sure can be improved.
Hi. As far as I understand from this line: https://github.com/open-mmlab/mmdetection3d/blob/7fec1d533bacf5b5c89d456f65c775c3cc458c72/mmdet3d/models/detectors/imvoxelnet.py#L47 ImVoxelNet only uses one feature map from the FPN's multi-scale feature maps. Am I miss-interpreting something? Can you enlighten me?