Workshop on Vision with Biased or Scarce Data

VBSD 2018


Computer Graphics Computer Vision & Pattern Recognition



The performance of vision algorithms on many key problems that were once considered hard is now astounding (e.g. object detection and categorization). In many cases however, significant amounts of data are necessary to achieve high performance. We would like to focus on problem domains where insufficient raw data is available for designing and training a system. We see scarcity of data in light of the task specific objective function, as well as the underlying distributions of the natural prior and the data sampling prior. As a concrete example from autonomous driving, one can generate a potentially infinite amount of simulated data for training by using game engines. This may nevertheless miss unlikely situations that have not been explicitly modeled (such as a paper bag being blown on the road: is it an obstacle worthy of action?) Sometimes it may even be difficult to just detect that the available data is scarce with respect to an objective function. With the ever-increasing pervasiveness of vision algorithms in the real-world this becomes an important point to consider.
The topic of scarce data has been and is continuously being covered in the ML and adjoining communities, for instance in transfer learning. From an ML perspective we feel that vision still has a special place when it comes to explaining observed data due to the vast amount of potential prior knowledge (e.g. the separation of appearance and geometry) and data generation models (e.g. computer graphics) available. Accordingly, we are looking for unique opportunities for computer vision in modeling data and algorithms under bias or scarcity. Towards this we pose the following questions:
When is data scarce or biased?
How can we detect bias or scarcity in data?
How can we improve and evaluate performance under bias or scarcity?
We solicit submissions that address any of the above questions as well as the following topics:
Modeling data and learning from inhomogeneous input
Structured and unstructured prior knowledge (e.g. language, multi-modality)
Combining domain-related and unrelated data and priors
Generative modeling for augmenting synthetic and real data
Converging rendering and inference (e.g. combine partial inference with partial synthesis)
Prior models for latent distributions that explain and generate images and videos
Automation of data and prior exploration with minimal human-in-the-loop
Design and evaluation of vision algorithms and systems under data scarcity
Detecting and quantifying bias and scarcity in data
Principled combinations of prior knowledge and learnt representations
Detecting and avoiding overfitting
Modeling and understanding generalization behavior with scarce data