Beyond MAP: Making Multiple Predictions: Diversity, DPPs and more.

Short Course (Tutorial), CVPR 2013
June 24, 2013
New: Videos of the talks are now available!

Overview


Computer vision systems must deal with a tremendous amount of uncertainty, from occlusion to varying appearance, lighting, and pose. Probabilistic models provide a principled framework for dealing with this uncertainty and for converting evidence from multiple noisy sources into a posterior belief about the world. Typically, an intelligent system will then use this belief to predict the most probable or maximum a-posteriori (MAP) hypothesis.

For a variety of reasons, a single prediction can be inadequate. If the model is misspecified, the training data are suboptimal, or complex and intractable learning objectives lead to significant optimization error, then the MAP solution may be unreliable. We might prefer to hedge our bets by making multiple predictions and then re-ranking or combining them to obtain a single answer.

This tutorial will cover models and techniques for generating multiple diverse predictions from structured probabilistic models:

Instructors


Dhruv Batra

Alex Kulesza

Dennis Park

Deva Ramanan


Schedule

Time          Topic Presenter Slides Video Link
9:00  -  9:30 Opening Remarks + Need for Multiple Diverse Solutions Dhruv Batra pdf pptx techtalks link
9:30  - 10:15 Multiple Solutions via M-Best MAP Dennis Park pdf key techtalks link (Skip to 23:30)
10:15 - 10:45 Coffee Break
10:45 - 11:30 Multiple Solutions via Diverse M-Best Dhruv Batra pdf pptx techtalks link
11:30 - 1:30 Lunch
1:30 - 2:00 Multiple Solutions via Sampling Deva Ramanan pdf key techtalks link
2:00 - 3:15 Multiple Solutions via DPPs Alex Kulesza pdf techtalks link (Skip to 39:13)
3:15 - 3:45 Coffee Break
3:45 - 4:30 DPPs (Continued) Alex Kulesza techtalks link
4:30 - 5:00 Closing Remarks + What can we do with diverse solutions? Dhruv Batra pdf pptx techtalks link (Skip to 44:30)

References and Resources

  • D. Park and D. Ramanan. N-best maximal decoders for part models. ICCV, 2011. [paper] [code]
  • D. Batra, P. Yadollahpour, A. Guzman-Rivera, and G. Shakhnarovich. Diverse M-Best Solutions in Markov Random Fields. ECCV, 2012. [paper]
  • A. Kulesza, B. Taskar. Determinantal Point Processes for Machine Learning. Foundations and Trends in Machine Learning, 5(2-3):123-286. [paper] [code]

© 2013 Virginia Tech