Method

AttentionTrack [AT]


Submitted on 25 Apr. 2022 15:38 by
Chuang Zhang (Tsinghua University)

Running time:0.03 s
Environment:1 core @ 2.5 Ghz (Python)

Method Description:
On the one hand, we utilizes attention module to
aggregate features from different down-sampling
rate, which can improve the robustness of feature
encoding to complex traffic environment and
various object size. On the other hand, we use
attention module to reprocess the features encoded
by backbone network to generate specific features
of detection and tracking tasks.
Parameters:
alpha=0.2
Latex Bibtex:

Detailed Results

From all 29 test sequences, our benchmark computes the HOTA tracking metrics (HOTA, DetA, AssA, DetRe, DetPr, AssRe, AssPr, LocA) [1] as well as the CLEARMOT, MT/PT/ML, identity switches, and fragmentation [2,3] metrics. The tables below show all of these metrics.


Benchmark HOTA DetA AssA DetRe DetPr AssRe AssPr LocA
CAR 66.62 % 66.77 % 67.24 % 72.19 % 79.91 % 71.66 % 83.00 % 84.29 %
PEDESTRIAN 40.85 % 39.86 % 42.19 % 44.44 % 66.96 % 47.91 % 69.25 % 78.37 %

Benchmark TP FP FN
CAR 29379 5013 1691
PEDESTRIAN 13325 9825 2039

Benchmark MOTA MOTP MODA IDSW sMOTA
CAR 79.77 % 82.34 % 80.51 % 254 64.68 %
PEDESTRIAN 47.18 % 74.42 % 48.75 % 364 32.45 %

Benchmark MT rate PT rate ML rate FRAG
CAR 68.15 % 25.69 % 6.15 % 457
PEDESTRIAN 29.55 % 41.58 % 28.87 % 778

Benchmark # Dets # Tracks
CAR 31070 869
PEDESTRIAN 15364 499

This table as LaTeX