TrctrD17: deepMOTTracktor17


Video not available.

Rendering of new sequences is currently deactivated due to heavy load.

Rendering of new sequences is currently deactivated due to heavy load.

Rendering of new sequences is currently deactivated due to heavy load.

Rendering of new sequences is currently deactivated due to heavy load.

Rendering of new sequences is currently deactivated due to heavy load.

Rendering of new sequences is currently deactivated due to heavy load.

Rendering of new sequences is currently deactivated due to heavy load.

Rendering of new sequences is currently deactivated due to heavy load.

Rendering of new sequences is currently deactivated due to heavy load.

Rendering of new sequences is currently deactivated due to heavy load.

Rendering of new sequences is currently deactivated due to heavy load.

Rendering of new sequences is currently deactivated due to heavy load.

Rendering of new sequences is currently deactivated due to heavy load.

Rendering of new sequences is currently deactivated due to heavy load.

Rendering of new sequences is currently deactivated due to heavy load.

Rendering of new sequences is currently deactivated due to heavy load.

Rendering of new sequences is currently deactivated due to heavy load.

Rendering of new sequences is currently deactivated due to heavy load.

Rendering of new sequences is currently deactivated due to heavy load.

Rendering of new sequences is currently deactivated due to heavy load.

Rendering of new sequences is currently deactivated due to heavy load.

Benchmark:

MOT17 |

Short name:

TrctrD17

Detector:

Public

Description:

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 to estimate 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.

Reference:

Y. Xu, A. Osep, Y. Ban, R. Horaud, L. Leal-Taixe, X. Alameda-Pineda. How To Train Your Deep Multi-Object Tracker. In , 2019.

Last submitted:

February 01, 2020 (4 years ago)

Published:

April 16, 2020 at 14:18:17 CET

Submissions:

1

Open source:

Yes

Hardware:

GTX TITAN XP

Runtime:

4.9 Hz

Benchmark performance:

Sequence MOTA IDF1 HOTA MT ML FP FN Rcll Prcn AssA DetA AssRe AssPr DetRe DetPr LocA FAF ID Sw. Frag
MOT1753.753.842.4458 (19.4)861 (36.6)11,731247,44756.196.442.742.546.873.444.676.680.40.71,947 (34.7)4,792 (85.4)

Detailed performance:

Sequence MOTA IDF1 HOTA MT ML FP FN Rcll Prcn AssA DetA AssRe AssPr DetRe DetPr LocA FAF ID Sw. Frag
MOT17-01-DPM37.237.831.3512883,92839.196.634.228.838.069.029.973.878.40.23681
MOT17-01-FRCNN35.936.431.25103643,73842.088.232.230.437.566.632.668.377.50.83594
MOT17-01-SDP38.537.031.76102473,67943.091.832.731.037.667.732.870.177.70.538101
MOT17-03-DPM65.659.046.451191,13634,67866.998.442.551.046.373.953.378.480.90.8196358
MOT17-03-FRCNN66.361.047.554201,10334,00967.598.544.451.248.473.953.578.080.60.7169331
MOT17-03-SDP69.862.749.161162,44928,91872.496.944.754.348.972.157.376.880.11.6207542
MOT17-06-DPM52.555.843.441892015,32354.897.044.142.952.469.844.979.482.50.275173
MOT17-06-FRCNN56.058.445.451604094,67560.394.644.846.353.169.749.277.181.60.3102261
MOT17-06-SDP56.258.645.657653954,67960.394.745.146.354.169.349.277.381.70.386244
MOT17-07-DPM40.745.635.26243429,59243.295.539.232.041.873.633.574.079.10.783268
MOT17-07-FRCNN39.644.634.57225549,56643.493.037.831.940.474.433.772.278.81.186268
MOT17-07-SDP41.545.935.47215779,21745.493.038.433.141.174.135.171.878.61.296337
MOT17-08-DPM27.131.528.673820415,13128.496.735.823.038.380.523.680.383.30.373121
MOT17-08-FRCNN27.131.528.793720015,12628.496.836.322.938.880.223.580.083.20.371108
MOT17-08-SDP28.732.529.3103624914,71030.496.335.624.238.578.424.979.182.60.495161
MOT17-12-DPM45.554.744.31544914,60146.997.852.237.656.577.739.081.383.50.12858
MOT17-12-FRCNN43.453.943.514471904,69445.895.452.336.357.176.738.079.183.50.22141
MOT17-12-SDP45.256.244.717442214,50048.195.052.838.156.877.240.078.983.20.23162
MOT17-14-DPM27.437.326.9108852112,79030.891.632.722.335.468.623.269.176.30.7100297
MOT17-14-FRCNN28.538.728.112801,08111,98135.285.732.225.036.064.126.664.975.41.4157429
MOT17-14-SDP28.738.728.513791,10911,91235.685.632.525.335.666.727.064.975.41.5162457

Raw data: