ECCV 2020 WSL tutorial: 5. Evaluating weakly-supervised methods скачать в хорошем качестве

ECCV 2020 WSL tutorial: 5. Evaluating weakly-supervised methods

37 minutes. Weakly-supervised learning evaluations are prone to error. We discuss why they are difficult, how to do them right, and how to move forward. We investigate what has gone wrong with the evaluation of many weakly-supervised learning tasks: test-set overfitting and implicit full supervision through validation. In our case study with weakly-supervised object localization (WSOL), we provide rectified evaluation benchmarks and protocols. We discover that six WSOL methods from 2016-2019 actually have not brought about the claimed improvements; in fact, the most basic CAM (CVPR 2016) is still the best object localizer. We link the difficulty of evaluation to the ill-posedness of weakly-supervised learning tasks and propose future directions for weakly-supervised learning. More videos from the same ECCV tutorial:    • ECCV 2020 tutorial on Weakly Supervised Le...   Tutorial website: https://hbilen.github.io/wsl-eccv20.g... Slides pdf: https://github.com/hbilen/wsl-eccv20....