Object Detection Models conventionally use Non-Maximum Suppression (NMS) as the default post-processing step to filter out redundant bounding boxes. However, this approach fails to efficiently give unified, averaged predictions across multiple models since they tend to remove less confident boxes having a significant overlap.
To mitigate this problem, we shall discuss an efficient pre-processing step called Weighted Boxes Fusion (WBF) that helps achieve a unified localized prediction across multiple detections.
We will discuss the working along with the PyTorch code for the WBF method in today's article.
No comments:
Post a Comment
Note: Only a member of this blog may post a comment.