search

XuyangBai / D3Feat

[TensorFlow] Official implementation of CVPR'20 oral paper - D3Feat: Joint Learning of Dense Detection and Description of 3D Local Features https://arxiv.org/abs/2003.03164
MIT License
259 stars 38 forks source link

Architecture issues #6

Closed tangbohu closed 4 years ago

tangbohu commented 4 years ago

Thanks for your great work! However, I am very curious about the sentense in the paper "All layers except the last one are followed by batch normalization and ReLU". I wonder why BN and ReLU are not adopted as in 2D CNNs. Looking forward to your kind reply!

XuyangBai commented 4 years ago

Hi @tangbohu Thanks for your interest. For the last layer, I follow the design of HardNet and L2Net in the 2D descriptor learning, where the ReLU is omitted. For BN, since we are learning a representation for each point, adding BN seems to hurt the representation power of the final descriptor (but some other works do add the bn for the last layer). And actually I have tried architectures both with and without BN and ReLU for the last layer, and this setting has quite a negligible effect on the final result if I remember correctly.