ICCV 2019 Poster Jiahe Li 2019 11 29

ICCV 2019, Poster Jiahe Li 2019. 11. 29 [MOT 15] Laura Leal-Taix´e, Anton Milan. Motchallenge 2015: Towards a enchmark for multi-target tracking. ar. Xiv, 2015. [MOT 16] Anton Milan, Laura Leal-Taix´e. Mot 16: A benchmark for multi-object tracking. ar. Xiv, 2016

A detector is all you need Faster R-CNN & Feature Pyramid Networks (FPN) Public Detector Tracktor [FPN] Lin et al. Feature pyramid networks for object detection. In CVPR, 2017.

Tracking extensions o Motion Model n Camera motion compensation (CMC) o Re-identification n Appearance vectors n Data association p Killed tracks p Newly detected tracks [CMC] Evangelidis et al. Parametric image alignment using enhanced correlation coefficient maximization. PAMI, 2008 [Re-ID] Ergys Ristani and Carlo Tomasi. Features for multi-target multi-camera tracking and re-identification. CVPR, 2018

Datasets o MOT 15 n 11 sequences n ACF o MOT 16 & MOT 17 n 7 sequences n DPM, FRCNN, SDP [ACF] Dollar et al. Fast feature pyramids for object detection. PAMI, 2014 [DPM] Pedro et al. Object detection with discriminatively trained part based models. PAMI, 2009. [FRCNN] Ren et al. Faster R-CNN: Towards real-time object detection with region proposal networks. NIPS, 2015. [SDP] Yang et al. Exploit all the layers: Fast and accurate cnn object detector with scale dependent pooling and cascaded rejection classifiers. CVPR, 2016.

Ablation Study The arrows indicate low or high optimal metric values.

MOTChallenge The arrows indicate low or high optimal metric values.

Object Visibility

Conclusion o Tracktor Q. If a detector can solve most of the tracking problems, what are the real situations where a dedicated tracking algorithm is necessary? A. Currently, the tracking problems are not solved even for Tracktor++. It is essential to develop the robust tracking algorithm. Most tracking algorithms, following the tracking-by-detection strategy, heavily relies on the performance of the underlying detection method.