Computer Vision News - August 2018

13 Visual Dialog 2018 Devi adds: “ Questions in Visual Dialog can have multiple plausible and/or similar (e.g. ' yeah ' and ' yes ') answers. Evaluation metrics till now, however, assumed a single ground truth answer. To improve this, we collected a dense set of human annotations for all 100 answer options for each test instance in VisDial . This makes evaluation much more robust! The winning entries of the Visual Dialog challenge will be picked based on these new dense human annotations. I am quite excited to see how the various submissions pan out with respect to this evaluation! ” She then concludes: “ A lot of the actual work is being done by the students leading this effort. It would be great if we can recognize them. All the organizers are listed here . Abhishek Das , Karan Desai and Ayush Shrivastava in particular have been doing a lot of the heavy lifting. ” Challenge Computer Vision News “This makes evaluation much more robust!”