Image list file format reqd for Object Detection Dataset


#1

I am working on object detection. I am trying to use my own data.
I don’t think I understand how best to get label information into a Dataset for object detection.

My list file looks like (3 sample lines):

|639|4|5|256|256|1|0.1914|0.2383|0.7852|0.7539|frame639.jpg|
|934|4|5|256|256|1|0.2148|0.2109|0.7734|0.7227|frame934.jpg|
|798|4|5|256|256|1|0.1484|0.1328|0.7891|0.75|frame798.jpg|

images are 256x256, class=1

My list file works with im2rec.py and creates a rec file.
This rec file appears to work with mx.image.ImageDetIter perfectly

However, I’m trying now to use more GluonCV functionality. My rec file doesn’t work correctly with:
ImageRecordDataset
It yields labels with shape = (9,) (not the same record as above, but) the data coming out of a ImageRecordDataset looks like:
label.shape: (9,) [4.0000000e+00 5.0000000e+00 2.5600000e+02 2.5600000e+02 0.0000000e+00
1.8750000e-01 1.2500000e-01 7.2656250e-01 8.0859375e-01]

Should I be using ImageRecordDataset?
Should I take a different approach to getting label information into a Dataset?
I assumed I wanted an IMageRecordDataset, i would use transfroms to create a DataLoader
But I’m stuck with an invalid dataset