Converting object detection to instance segmentation

[Carl_James]

Hi,

i’ve exported a dataset created with object detection then imported that dataset into an instance segmentation model so i can make the results more defined.

When I test the new segmentation model it works gr8 however the result of the inference shows the bounding box instead of the human we are detecting.

Is this because the dataset was originally created with an object detection model?

Here’s the result:

Is there another process I should go through to fix this?

Thanks

[Jacob_Witt]

Hi @Carl_James - you need to label the players with polygons in order to get polygon outputs. Otherwise the instance segmentation model treats each bounding box as a polygon in itself.

[Carl_James]

Thanks Jacob,

Will i need to annotate each image manually again with the polygons bound to the players? or is there a quick way to do this for my dataset that has thousands of images.

thanks

[Jacob_Witt]

You can go image-by-image and select Convert Boxes to Smart Polygons, though there is no way to do this action in bulk.

Is there a particular reason you need instance segmentation?