Skip to first unread message

ankita raj

unread,
Mar 10, 2018, 4:17:41 AM3/10/18
to Caffe Users
Hi,

I am trying to feed 6 channel images to my caffe SSD network. From what I have discovered so far, I must do this using lmdb since the annotated data layer in ssd only supports lmdb (correct me if I am wrong). I have managed to create a 6-channel cv::Mat in convert_annoset.cpp, and saved it as lmdb (I had to do this without encoding, since cv::imencode does not support more than 4 channels).

Now, I can successfully read the created lmdb data using python, and it has the correct dimensions. However, when I pass this to the caffe network, it throws the following error:

data_transformer.cpp: Check failed: datum_height == height

It turns out that the datum height and width do not match the expected blob height. In fact, the datum dimensions change randomly every time I try to load the network. From what I could figure out, this is because data_transfomer.cpp only supports encoded datum. But in my case, I cannot use encoded datum because of the limitations of cv::imencode.

Could someone please help me around with this ?
Reply all
Reply to author
Forward
0 new messages