Alan Lukežič presents DCF Tracking

On 2020-07-09 11:00:00 at G205, Karlovo náměstí 13, Praha 2
Template-based discriminative trackers, in particular the discriminative
correlation filters (DCF), have been the dominant tracking methodology for the
last five years. DCFs represent the target as a rectangular region, which often
leads to drift and failure when tracking non-compact objects. A more accurate
representation is per-pixel segmentation, but segmentation alone is prone to
failure in presence of similar objects.

In the first part of the talk we will present an improvement of the standard
DCF formulation using an approximate segmentation (CSRDCF), which achieves a
real-time performance on a CPU. A major drawback of this method is the use of
hand-crafted algorithms and features.

In the second part of the talk, we will thus present our recent discriminative
single-shot segmentation tracker (D3S), which merges a DCF and segmentation
within a single-stage deep neural network. Trained only for segmentation as the
primary output, and without per-dataset finetuning, D3S achieves a
state-of-the-art performance on major tracking benchmarks.

Link to the D3S paper:

Link to the D3S video:
Za obsah zodpovídá: Petr Pošík