Pretrained model 'voxel_regre' requires 3d input instead of 2d

[xiangyu]

When running with model 'voxel_regre':

python -m camera.capture --data_root=$HOME/data --out_root=$HOME/data/univue/output --model_name=voxel_regre

Error reported:

File "/web/research/vendors/depth-hand/code/camera/capture.py", line 214, in update
depth_image, cube, sess, ops)
File "/web/research/vendors/depth-hand/code/camera/capture.py", line 185, in detect_region
feed_dict=feed_dict)
File "/home/qwe/.local/lib/python2.7/site-packages/tensorflow/python/client/session.py", line 900, in run
run_metadata_ptr)
File "/home/qwe/.local/lib/python2.7/site-packages/tensorflow/python/client/session.py", line 1111, in _run
str(subfeed_t.get_shape())))
ValueError: Cannot feed value of shape (1, 128, 128, 1) for Tensor u'Placeholder:0', which has shape '(1, 64, 64, 64, 1)'

It says the model uses 3d input, but we feed 2d input.

[xkunwu]

Hi, thanks for your feedback. "voxel_regre" is indeed a 3d model and requires its own preprocessed data (should present in the "data_root"). Unfortunately I cannot upload that data due to the huge size.

But you can actually run the estimation part first, which will produce the required data, then you can run the realtime capture part without problem. Be aware that it may take some time for the preprocessing and training, and the data is really huge.