[R] Data-centric AI development approach gives us 5,5-8% mAP on PASCAL VOC 2012 – June 28, 2022 at 05:36AM

How important is clean data for how your AI models perform?

According to our initial experiments – very important. Two people improved the primary model metric by 13% in a week using state-of-the-art confidence learning to clean up PASCAL.

In the next iteration, we explore the field a bit deeper, trying to avoid controversy. SPOILER: we still get a nice mAP metric boost.

To learn more about the follow-up results and the plans for the future, check out our article: https://hasty.ai/content-hub/articles/further-pascal-voc-2012-exploration-and-plans-for-the-future

​

Discussion of the initial results:

• https://www.reddit.com/r/MachineLearning/comments/uc9z2y/p_we_cleaned_up_pascal_and_improved_map_by_13/ia6caqg/?context=3
• https://www.reddit.com/r/computervision/comments/uc9x8t/we_cleaned_up_pascal_and_improved_map_by_13/i69iyii/?utm_source=share&utm_medium=web2x&context=3

​

Disclaimer: We used our own platform to clean up the data and the article, therefore, contains self-promotion. However, the article mainly focuses on the results we achieved.