Deep Network Flow for Multi-Object Tracking: Supplemental Material

The supplemental material of our deep network flow approach for multi-object tracking contains the following items: • Details on the formulation of deep network flows (Section 1) • An on-line version of the tracker (Section 2) • Qualitative results (Section 3) 1. Details on the formulation of deep network flows First, we want to provide further details of our formulation of deep network flows and for computing gradients of the linear programming solution w.r.t. the cost functions. 1.1. The nullspace of C is larger than the trivial solution 0 In Section 3.2.1 of the main paper, we smooth the lower level problem, i.e., the linear program. We get rid of the box constraints with log-barriers and remove the flow conservation constraints (matrix C) with a change of basis, which requires the null space of C. The matrix C ∈ R2K×M models the flow conservation constraints xin i + ∑