How to Train Your Deep Multi-Object Tracker
Open accessThe DOI matches its Crossref record; 24 of the 24 references it cites were found in a registry or an index.
Abstract
The recent trend in vision-based multi-object tracking (MOT) is heading towards leveraging the representational power of deep learning to jointly learn to detect and track objects. However, existing methods train only certain sub-modules using loss functions that often do not correlate with established tracking evaluation measures such as Multi-Object Tracking Accuracy (MOTA) and Precision (MOTP). As these measures are not differentiable, the choice of appropriate loss functions for end-to-end training of multi-object tracking methods is still an open research problem. In this paper, we bridge this gap by proposing a differentiable proxy of MOTA and MOTP, which we combine in a loss function suitable for end-to-end training of deep multi-object trackers. As a key ingredient, we propose a Deep Hungarian Net (DHN) module that approximates the Hungarian matching algorithm. DHN allows estimating the correspondence between object tracks and ground truth objects to compute differentiable proxies of MOTA and MOTP, which are in turn used to optimize deep trackers directly. We experimentally demonstrate that the proposed differentiable framework improves the performance of existing multi-object trackers, and we establish a new state of the art on the MOTChallenge benchmark. Our code is publicly available from https://github.com/yihongXU/deepMOT.
Overview
The claims and statistics of this paper haven’t been checked yet.
Free: it reads the open-access text and checks the first 5 claims that cite a source we can check. Create a free account to check every claim.
Integrity signals
Registry
Cites
Statistics
Verification
Claims the paper makes
The claims and statistics of this paper haven’t been checked yet.
Free: it reads the open-access text and checks the first 5 claims that cite a source we can check. Create a free account to check every claim.
Self-check
Percentages, sums and reported differences re-computed from the paper’s own numbers.
The self-check needs the paper’s full text.
Citations
Cites
Hochreiter, S., & Schmidhuber, J. (1997) · Neural Computation · doi:10.1162/neco.1997.9.8.1735 · Cited by 102,318 · cited 1×
“The recurrent unit of the two recurrent architectures, seq and paral, is also ablated between long-short term memory units (lstm) [23] and gated recurrent units (gru) [10].”
BackgroundFaster R-CNN: Towards Real-Time Object Detection with Region Proposal Networks
Ren, S., He, K., Girshick, R., & Sun, J. (2017) · IEEE Transactions on Pattern Analysis and Machine Intelligence · doi:10.1109/tpami.2016.2577031 · Cited by 73,472
Histograms of Oriented Gradients for Human Detection
Dalal, N., & Triggs, B. (2005) · 2005 IEEE Computer Society Conference on Computer Vision and Pattern Recognition (CVPR'05) · doi:10.1109/cvpr.2005.177 · Cited by 33,584
Focal Loss for Dense Object Detection
Lin, T.-Y., Goyal, P., Girshick, R., He, K., & Dollar, P. (2017) · 2017 IEEE International Conference on Computer Vision (ICCV) · doi:10.1109/iccv.2017.324 · Cited by 30,581
FaceNet: A unified embedding for face recognition and clustering
Schroff, F., Kalenichenko, D., & Philbin, J. (2015) · 2015 IEEE Conference on Computer Vision and Pattern Recognition (CVPR) · doi:10.1109/cvpr.2015.7298682 · Cited by 15,267 · cited 4×
“This external ReID module is a feature extractor network with a ResNet-50 backbone, trained using a triplet loss [44] on the MOTChallenge video sequences.”
BackgroundMethodInfluentialThe Hungarian method for the assignment problem
Kuhn, H. W. (1955) · Naval Research Logistics Quarterly · doi:10.1002/nav.3800020109 · Cited by 13,120
Dimensionality Reduction by Learning an Invariant Mapping
Hadsell, R., Chopra, S., & LeCun, Y. (2006) · 2006 IEEE Computer Society Conference on Computer Vision and Pattern Recognition - Volume 2 (CVPR'06) · doi:10.1109/cvpr.2006.100 · Cited by 5,336
An algorithm for tracking multiple targets
Reid, D. (1979) · IEEE Transactions on Automatic Control · doi:10.1109/tac.1979.1102177 · Cited by 3,059
High Performance Visual Tracking with Siamese Region Proposal Network
Li, B., Yan, J., Wu, W., Zhu, Z., & Hu, X. (2018) · 2018 IEEE/CVF Conference on Computer Vision and Pattern Recognition · doi:10.1109/cvpr.2018.00935 · Cited by 3,023 · cited 2×
“3, we outline the results of the two MOT-by-SOT baselines (using GOTURN [18] and SiamRPN [29]).”
MethodEvaluating Multiple Object Tracking Performance: The CLEAR MOT Metrics
Bernardin, K., & Stiefelhagen, R. (2008) · EURASIP Journal on Image and Video Processing · doi:10.1155/2008/246309 · Cited by 2,918
MOT16: A Benchmark for Multi-Object Tracking
Milan, A., Leal-Taixe, L., Reid, I., Roth, S., & Schindler, K. (2016) · arXiv · doi:10.48550/arxiv.1603.00831 · Cited by 1,309 · cited 2×
“We generate distance matrices D using ground-truth bounding boxes and public detections, provided by the MOT challenge datasets [34, 27].”
BackgroundMethodLearning to Track at 100 FPS with Deep Regression Networks
Held, D., Thrun, S., & Savarese, S. (2016) · Lecture Notes in Computer Science · doi:10.1007/978-3-319-46448-0_45 · Cited by 1,199
Tracking Without Bells and Whistles
Bergmann, P., Meinhardt, T., & Leal-Taixe, L. (2019) · 2019 IEEE/CVF International Conference on Computer Vision (ICCV) · doi:10.1109/iccv.2019.00103 · Cited by 1,110
Joint Unsupervised Learning of Deep Representations and Image Clusters
Yang, J., Parikh, D., & Batra, D. (2016) · 2016 IEEE Conference on Computer Vision and Pattern Recognition (CVPR) · doi:10.1109/cvpr.2016.556 · Cited by 879 · cited 3×
“For MOT16 and MOT17, we average the results obtained using the three sets of provided public detections (DPM [18], SDP [16] and Faster R-CNN [43]).”
MethodGlobally-optimal greedy algorithms for tracking a variable number of objects
Pirsiavash, H., Ramanan, D., & Fowlkes, C. C. (2011) · CVPR 2011 · doi:10.1109/cvpr.2011.5995604 · Cited by 790
High-Speed tracking-by-detection without using image information
Bochinski, E., Eiselein, V., & Sikora, T. (2017) · 2017 14th IEEE International Conference on Advanced Video and Signal Based Surveillance (AVSS) · doi:10.1109/avss.2017.8078516 · Cited by 707 · cited 1×
“here is calculated with IoU, it is thus the higher the better, some for the following evaluations. 7 Mode Method MOTA(") MOTP(") IDF1(") MT(") ML(#) FP(#) FN(#) IDSw(#) Offline IOU [7] 45.5 76.9 39.4 15.7% 40.5% 19,993 281,643 5,988 MHT bLSTM [19] 47.5 77.5 51.9 18.2% 41.7% 25,981 268,042 2,069 EDMT…”
InfluentialMultiple Hypothesis Tracking Revisited
Kim, C., Li, F., Ciptadi, A., & Rehg, J. M. (2015) · 2015 IEEE International Conference on Computer Vision (ICCV) · doi:10.1109/iccv.2015.533 · Cited by 666 · cited 1×
“[22] leverages pre-learned CNN features or a bilinear LSTM [23] to learn the longterm appearance model.”
BackgroundMOTS: Multi-Object Tracking and Segmentation
Voigtlaender, P., Krause, M., Osep, A., Luiten, J., Sekar, B. B. G., Geiger, A., & Leibe, B. (2019) · 2019 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) · doi:10.1109/cvpr.2019.00813 · Cited by 658 · cited 3×
“Other methods [26, 51, 47] learn identity-preserving embeddings for data association using deep neural networks, trained using contrastive [17], triplet [44] or quadruplet loss [47].”
BackgroundMethodYang, F., Choi, W., & Lin, Y. (2016) · 2016 IEEE Conference on Computer Vision and Pattern Recognition (CVPR) · doi:10.1109/cvpr.2016.234 · Cited by 608
Chen, L., Ai, H., Zhuang, Z., & Shang, C. (2018) · 2018 IEEE International Conference on Multimedia and Expo (ICME) · doi:10.1109/icme.2018.8486597 · Cited by 530 · cited 2×
“0 25937 247822 1802 MOTDT17 [29] 50.”
Learning by Tracking: Siamese CNN for Robust Target Association
Leal-Taixe, L., Canton-Ferrer, C., & Schindler, K. (2016) · 2016 IEEE Conference on Computer Vision and Pattern Recognition Workshops (CVPRW) · doi:10.1109/cvprw.2016.59 · Cited by 444 · cited 2×
“Other methods [26, 51, 47] learn identity-preserving embeddings for data association using deep neural networks, trained using contrastive [17], triplet [44] or quadruplet loss [47].”
BackgroundMethodChu, Q., Ouyang, W., Li, H., Wang, X., Liu, B., & Yu, N. (2017) · 2017 IEEE International Conference on Computer Vision (ICCV) · doi:10.1109/iccv.2017.518 · Cited by 406
Online Multi-Object Tracking with Dual Matching Attention Networks
Zhu, J., Yang, H., Liu, N., Kim, M., Zhang, W., & Yang, M.-H. (2018) · Lecture Notes in Computer Science · doi:10.1007/978-3-030-01228-1_23 · Cited by 377 · cited 2×
“preserving embeddings for data association [24, 43, 46], learning the appearance model of individual targets [12, 50] and learning to regress the pose of the detected targets [4].”
BackgroundChu, P., & Ling, H. (2019) · 2019 IEEE/CVF International Conference on Computer Vision (ICCV) · doi:10.1109/iccv.2019.00627 · Cited by 246
Cited by
StrongSORT: Make DeepSORT Great Again
Du, Y., Zhao, Z., Song, Y., Zhao, Y., Su, F., Gong, T., & Meng, H. (2023) · IEEE Transactions on Multimedia · doi:10.1109/tmm.2023.3240881 · Cited by 1,191
“Recently, several joint tracking methods [57, 59, 28, 51] have been proposed to jointly train detection and other components, such as motion, embedding and association models.”
MethodMOTChallenge: A Benchmark for Single-Camera Multiple Target Tracking
Dendorfer, P., Os̆ep, A., Milan, A., Schindler, K., Cremers, D., Reid, I., Roth, S., & Leal-Taixé, L. (2021) · International Journal of Computer Vision · doi:10.1007/s11263-020-01393-0 · Cited by 374
“To address this discrepancy DeepMOT (Xu et al. 2020) proposes the missing link—a differentiable matching layer that allows expressing a soft, differentiable variant of MOTA and MOTP.”
BackgroundMethodSelf-citationTransMOT: Spatial-Temporal Graph Transformer for Multiple Object Tracking
Chu, P., Wang, J., You, Q., Ling, H., & Liu, Z. (2023) · 2023 IEEE/CVF Winter Conference on Applications of Computer Vision (WACV) · doi:10.1109/wacv56688.2023.00485 · Cited by 258
Improving Multiple Pedestrian Tracking by Track Management and Occlusion Handling
Stadler, D., & Beyerer, J. (2021) · 2021 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) · doi:10.1109/cvpr46437.2021.01081 · Cited by 147
“Method MOTA IDF1 FP FN IDSW PV [19] 50.4 50.8 2600 86780 1061 Tracktor++ [2] 54.4 52.5 3280 79149 682 DeepMOT [35] 54.8 53.4 2955 78765 645 GSM [22] 57.0 58.2 4332 73573 475 TMOH ( online ) 63.2 63.5 3122 63376 635 TPM [25] 51.3 47.9 2701 85504 569 MLT [37] 52.8 62.6 5362 80444 299 MPNTrack [6]…”
MethodInfluentialA General Recurrent Tracking Framework without Real Data
Wang, S., Sheng, H., Zhang, Y., Wu, Y., & Xiong, Z. (2021) · 2021 IEEE/CVF International Conference on Computer Vision (ICCV) · doi:10.1109/iccv48922.2021.01297 · Cited by 70
“Most famous MOT methods [9, 5, 30, 17, 24] belong to the tracking-by-detection paradigm, which heavily depends on the performance of the underlying detector.”
Method
Metadata
- Authors
- Xu, Y., sep, A., Ban, Y., Horaud, R., Leal-Taixe, L., & Alameda-Pineda, X.
- Venue
- 2020 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR)
- Publisher
- IEEE
- Published
- 2020-06
- Pages
- 6786-6795
- ISSN
- 1063-6919 (ISSN-L)
- OpenAlex
- W3034679090
- Cited by
- 232
- References
- 24
- FWCI
- 13.30
- Type
- conference paper
- Language
- en
- Open access
- green
About these checks
Registry, retraction, Cites and statistics results are computed, not judged, with one exception: a printed reference that no index lists is read once by a model before it counts as a citation error. Claims are extracted by a model and kept only when their quote appears verbatim in the text; each cited source is then checked, with a model’s help, for whether it says what the claim relies on. Those verdicts can be wrong — read the claim’s quote and the cited source before relying on one.
Statistics and arithmetic are recomputed, not judged.
Confidence bands are shown only for fields calibrated on our annotated test set. None is calibrated yet, so no band is shown.
What counts as a claim ›How we measure confidence ›How we check this ›
Are you an author, or do you see a problem with this page? Request a review
Data sources
Bibliographic data: Semantic Scholar, OpenAlex, Crossref.
Contains information from the Semantic Scholar Academic Graph Datasets, made available under the ODC Attribution License (ODC-BY 1.0).
Semantic Scholar’s data is described in: Kinney, R., et al. (2023). The Semantic Scholar Open Data Platform. arXiv:2301.10140.