Done
classes_dict.json- Use a json file to read classes, it should be easier to rerun the model if we are training with new scheme, which has more data.
VocDatasetUse recurssive parsing function
Add
areakey totargetfor future calculating UoIAdd
iscrowdkey totargetto indicated overlapping objects
Datadirectory tree- To make dataset instancing easier
Todos
Recent goals
Rewrite xml modification function
W/H of the image –> add a get_height/width function in
VocDatasetOverlapping information as in
difficultReprocess the xmls
Make sure the boxes are more accurate
What to do if “no object” show up when getting the item
Remove the rgb in the images
Future goals
Implement the rest of the FasterRcnn
Look into TEED
Any changes can be made to the receptive field?
Optimization?