You signed in with another tab or window. Reload to refresh your session.You signed out in another tab or window. Reload to refresh your session.You switched accounts on another tab or window. Reload to refresh your session.Dismiss alert
I'm using transfer learning for training an object detection model, the model convered 3 classes:
people
bicycle
custom sign
the people and bicycle will get from public dataset like PASCAL VOC, which the amount is big, while the custom sign is a private dataset with very small amount, say 1000 images with a single label in each.
As the training tool requires a balance dataset for each class for achieve a good accuracy, so I'm planning import the full PASCAL VOC and custom dataset into 2 CVAT tasks (2 tasks are in one project), after labeled the custom sign data, when doing the export, I expect the exported (I prefer the format KITTI) dataset can achieve:
limited classes
as PASCAL VOC has 20 classes, here I only want export people, bicycle, and plus custom sign
Balance distribution for that 3 classes
as PASCAL VOC has thousands data for each class, for balancing with custom sign, each class need to align with amount 1000.
does the tool now support this?
The text was updated successfully, but these errors were encountered:
shaojun
changed the title
Support export balance dataset from imbalance source(project)
Support export balance dataset from imbalance source
Jan 4, 2022
I use datumaro (https://github.com/openvinotoolkit/datumaro) for this usecase. It used to be a part of cvat and it is now it's own thing. You can download the tasks in one format, convert it to the format you want and remove classes you don't like. Balancing can be easily achieved by just sampling the annotations and removing the rest.
There is no command which does exactly want you want. But you can write a small script to achieve your goal. In the issue I linked there are some suggestions on how to get it working.
It's also possible to not use datumaro or other libraries like that and write the code for it completly yourself.
Hi, thanks for the great effort.
I'm using transfer learning for training an object detection model, the model convered 3 classes:
the
people
andbicycle
will get from public dataset like PASCAL VOC, which the amount is big, while thecustom sign
is a private dataset with very small amount, say 1000 images with a single label in each.As the training tool requires a balance dataset for each class for achieve a good accuracy, so I'm planning import the full
PASCAL VOC
andcustom
dataset into 2 CVAT tasks (2 tasks are in one project), after labeled thecustom sign
data, when doing the export, I expect the exported (I prefer the formatKITTI
) dataset can achieve:as
PASCAL VOC
has 20 classes, here I only want exportpeople
,bicycle
, and pluscustom sign
as
PASCAL VOC
has thousands data for each class, for balancing withcustom sign
, each class need to align with amount 1000.does the tool now support this?
The text was updated successfully, but these errors were encountered: