Data Format for Multi-label/ Bounding Box

[Travis]

Hi,

I have just started using Caffe, and I am trying to solve the following problem:

My input file looks like:

Image_filename_1  x1, y1, x2, y2

Image_filename_2  x1, y1, x2, y2

where Image_filename is the name of my image file. All the images belong to the same category, eg: cats.

x1,y1,x2,y2 are the co-ordinates of the bounding box of the image normalized between 0-1.

My aim is to predict the bounding box using Caffe and then to calculate the EuclideanLoss in the output.

The goal is to output the bounding box co-ordinates for new images.

Question:

What will be the best data format to be used? lmdb or hdf5?

• Is this possible with lmdb format? If yes, will the 4 co-ordinates be treated as 4 labels? Can we convert the above text file with convert_imageset.cpp?
• How can I convert my above file and image folder into hdf5 format?