Eliminating Exposure Bias and Metric Mismatch in Multiple Object Tracking

Abstract

Identity Switching remains one of the main difficulties Multiple Object Tracking (MOT) algorithms have to deal with. Many state-of-the-art approaches now use sequence models to solve this problem but their training can be affected by biases that decrease their efficiency. In this paper, we introduce a new training procedure that confronts the algorithm to its own mistakes while explicitly attempting to minimize the number of switches, which results in better training.We propose an iterative scheme of building a rich training set and using it to learn a scoring function that is an explicit proxy for the target tracking metric. Whether using only simple geometric features or more sophisticated ones that also take appearance into account, our approach outperforms the state-of-the-art on several MOT benchmarks.

Long-Term Tracking With Deep Tracklet Association

“…In [39], LSTM was adopted to improve the performance of appearance modeling. Maksai et al [40] introduced an iterative scheme to minimize the number of identity switches during training and learned a scoring function for association. Chen et al [41] presented a method to align appearance features of tracklets.…”

HOTA: A Higher Order Metric for Evaluating Multi-object Tracking

“…This became the CLEAR MOT metrics (Bernardin and Stiefelhagen 2008) which positions the MOTA metric as the main metric for tracking evaluation alongside other metrics such as MOTP. MOTA was adopted for evaluation in the PETS workshop series (Ellis and Ferryman 2010) and remains, to this day, the most commonly used metric for evaluating MOT algorithms, although it has often been highly criticised (Shitrit et al 2011;Bento and Zhu 2016;Leichter and Krupka 2013;Leal-Taixé et al 2017;Milan et al 2013;Ristani et al 2016;Dave et al 2020;Luiten et al 2020;Maksai and Fua 2019;Wang et al 2019;Maksai et al 2017;Yu et al 2016;Dendorfer et al 2020;Luo et al 2014) for its bias toward overemphasizing detection over association (see Fig. 1), as well as a number of other issues (see Sect.…”

“…The IDF1 metric (Ristani et al 2016) which was proposed specifically for tracking objects through multiple cameras has been used by 'multi-camera MOT' benchmarks such as Duke-MTMC (Ristani et al 2016), AI City Challenge (Naphade et al 2017) and LIMA (Layne et al 2017). IDF1 has also recently been implemented as a secondary metric on the MOTChallenge benchmark, and has become preferred over MOTA for evaluation by a number of single camera tracking methods (Maksai and Fua 2019;Maksai et al 2017;Wang et al 2019) due to its focus on measuring association accuracy over detection accuracy (see Fig.1). IDF1 however exhibits unintuitive and nonmonotonic behaviour in regards to detection (see Sect.…”

“…In recent years IDF1 has been adopted by a number of MOT papers (Maksai and Fua 2019;Maksai et al 2017;Wang et al 2019) for MOT evaluation instead of MOTA, as these papers wish to adequately account for association in evaluation, which is lacking from MOTA. However, IDF1 produces counter-intuitive and non-monotonic results for measuring detection.…”

“…We restrict our evaluation to only those methods that are published in peer reviewed journals and conferences. We evaluate 37 different trackers (Brasó and Leal-Taixé 2020;Wang et al 2019;Bergmann et al 2019;Sheng et al 2018;Maksai and Fua 2019;Yoon et al 2020;Zhu et al 2018;Keuper et al 2018;Chen et al 2017Chen et al , 2019Xu et al 2019;Henschel et al 2018Henschel et al , 2019Long et al 2018;Kim et al 2015Kim et al , 2018Yoon et al 2018;Fu et al 2018Fu et al , 2019Chu and Ling 2019;Liu et al 2019;Song et al 2019;Karunasekera et al 2019;Babaee et al 2019;Cavallaro 2016, 2019;Bewley et al 2016;Bochinski et al 2017;Baisa 2018;Song and Jeon 2016;Baisa 2019;Eiselein et al 2012;Kutschbach et al 2017; Baisa and Wallace 2019) on MOT17 (Milan et al 2016). This includes all of the trackers for which the relevant bibliographic information was available when this analysis was performed on the 1st April 2020.…”