Object Detection on Redacted Documents

This is another (not so) crazy example where using simple Fastai & IceVision defaults gets you incredible results: 73% on the COCO metric on brand new object detection dataset of redacted documents.

To give you an idea on how good is 73%, Faster-RCNN with a ResNet101 has a 36% on the COCO dataset. The latter is obviously harder but Alex’ results are still very impressive.

Work done by the Alex @strickvl

1 Like