BBT: Bounding box based tracking enhanced by track-projection re-identification

MOT20-06


Benchmark:

MOT20 | MOT17 |

Short name:

BBT

Detector:

Public

Description:

Tracker uses only bounding boxes, no visual information.
Bounding boxes are associated frame-to-frame maximizing the IOU metric.
Re-identification phase projects discovered tracks to past/future tracks within a time window of 50 frames and quantifies the track matching score.
Fragmented tracks with high matching score are linked together, missing bounding boxes interpolated.
The proposed tracker is extremely fast compared to visual-based methods, and more accurate than (similar-to-our-work method) IOU17 thanks to the re-identification phase,
The proposed tracker is used to discover inter-person relationships in videos.

Reference:

Anonymous submission

Last submitted:

May 21, 2020 (2 months ago)

Published:

May 21, 2020 at 16:57:06 CET

Submissions:

2

Project page / code:

n/a

Open source:

No

Hardware:

Intel(R) Xeon(R) CPU E5-2620, 1 core

Runtime:

8.0 Hz

Benchmark performance:

Sequence MOTA IDF1 MOTP MT ML FP FN Recall Precision FAF ID Sw. Frag
MOT2046.842.278.0312 (25.1)289 (23.3)35,014236,17654.488.97.83,880 (71.4)7,207 (132.6)

Detailed performance:

Sequence MOTA IDF1 MOTP MT ML FP FN Recall Precision FAF ID Sw. Frag
MOT20-0466.452.080.6231704,25186,02768.697.82.01,7323,109
MOT20-0625.128.372.43110715,69582,57337.876.215.61,2282,295
MOT20-0752.147.276.836191,35214,24557.093.32.3249382
MOT20-0812.623.870.5149313,71653,33131.263.817.06711,421

Raw data: