Joachim Denzler and Erik Rodner
Mon 25 Apr 2016, 15:00 - 16:00

If you have a question about this talk, please contact: Steph Smith (ssmith32)

In biodiversity, research sensors already collect numerous data that is
difficult if not impossible to evaluate by hand. Examples are camera traps
continuously monitoring the environment to evaluate distribution of species
over a certain area and time. This leads to a big-data problem where the
data contains the knowledge the researcher is searching for. Computer vision
and machine learning can provide methods for (semi-)automatic evaluation and
structuring of large data streams. Currently, most of the research in this
area is driven by the recently founded Michael Stifel Center for Data Driven
and Simulation Science, Jena (

In the first part of the presentation, we point to potentials of life-long
learning, especially active and incremental learning in such a monitoring
application. Individual aspects, i.e. fine-grained recognition, recent
methods for active learning, and novelty detection, are presented.
Experimental evaluation (automatic bird species and moth classification,
incremental evaluation of images from camera traps in nature) summarizes
first results for automatic analysis of images and videos that keep the
human in the loop. Finally, a method for interactive image retrieval to
support researchers from biology is demonstrated.

In the second part of the talk, our work on visual fine-grained recognition
will be presented in detail. Our general-purpose techniques allow for
categorizing very similar looking species by learning the tiny details that
matter for discrimination. In particular, we developed an unsupervised
learning algorithm that re-focuses the attention of pre-trained deep neural
networks towards constellations of image regions likely related to object
parts. This allowed us to achieve state-of-the-art performance on several
datasets (flowers, dog breeds, etc...) including up to 76% for
distinguishing between 555 bird categories.