search

YoungXIAO13 / PoseFromShape

(BMVC 2019) PyTorch implementation of Paper "Pose from Shape: Deep Pose Estimation for Arbitrary 3D Objects"
http://imagine.enpc.fr/~xiaoy/PoseFromShape/
MIT License
170 stars 34 forks source link

shape and image feature dimension #12

Closed kartikgupta-at-anu closed 4 years ago

kartikgupta-at-anu commented 4 years ago

can you please tell me what are the specific shape and image feature dimensions used for the results in the paper, for different datasets? also can you indicate the training time for each dataset?

YoungXIAO13 commented 4 years ago

For all the datasets used in the paper, the image feature dimension is 1024 from the last layer of ResNet-18, and the shape feature dimension is 256 for one rendered image and 1024 for pointcloud.

One thing to notice is that there are 12 rendered images but only one pointcloud used to represent each shape.

As for the training time, report on one Titan-X GPU with 4 workers: Pascal3D: ~ 3 days. ObjectNet3D: ~2 days. ShapeNet: ~ 1 week. You can accelerate the training by loading the data from a SSD if allowed.