Data association problems are an important component of many computer visionapplications, with multi-object tracking being one of the most prominentexamples. A typical approach to data association involves finding a graphmatching or network flow that minimizes a sum of pairwise association costs,which are often either hand-crafted or learned as linear functions of fixedfeatures. In this work, we demonstrate that it is possible to learn featuresfor network-flow-based data association via backpropagation, by expressing theoptimum of a smoothed network flow problem as a differentiable function of thepairwise association costs. We apply this approach to multi-object trackingwith a network flow formulation. Our experiments demonstrate that we are ableto successfully learn all cost functions for the association problem in anend-to-end fashion, which outperform hand-crafted costs in all settings. Theintegration and combination of various sources of inputs becomes easy and thecost functions can be learned entirely from data, alleviating tedioushand-designing of costs.
展开▼
