Lifted Disjoint Paths with Application in Multiple Object Tracking

Authors: Andrea Hornakova, Roberto Henschel, Bodo Rosenhahn, Paul Swoboda

ICML 2020 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental We conduct several experiments on MOT showing the merit of using lifted disjoint paths for the tracking problem.
Researcher Affiliation Academia 1Computer Vision and Machine Learning, Max Planck Institute for Informatics, Saarbrücken, Saarland, Germany 2Institut for Image Processing, Leibniz University Hannover, Hannover, Niedersachsen, Germany.
Pseudocode Yes Algorithm 1 Separation for lifted path inequalities (10) ... Algorithm 2 Separation for lifted path-induced cut inequalities (11) and (12) ... Algorithm 3 Extract_path(P 1, v, w)
Open Source Code Yes Our code is available at https://github.com/ Andrea Hor/Lif T_Solver.
Open Datasets Yes We conduct extensive experiments on three challenging benchmarks: MOT15 (Leal Taixé et al., 2015), MOT16 and MOT17 (Milan et al., 2016), resulting in 39 test sequences.
Dataset Splits Yes We perform analysis and parameter tuning for our tracker on the MOT17 train set, even when our tracker is applied to the MOT15 sequences to ensure that our tracker is not prone to overfitting. We follow the MOT challenge protocol and use the detections provided by the respective benchmarks. All experiments on the training set are evaluated using a leave-one-out crossvalidation.
Hardware Specification No No specific hardware details (e.g., CPU, GPU models, or cloud computing instance types) are mentioned for the experimental setup within the provided text.
Software Dependencies Yes We solve the lifted disjoint paths problem (3) with the state of the art integer linear program solver Gurobi (Gurobi Optimization, 2019).
Experiment Setup Yes Input detections are rejected if Tracktor s detector outputs a confidence score σactive 0.5. Tracktor also applies a non-maxima-surpression on the reshaped input detections, where we use the threshold λnew = 0.6.