Method

Object Permanence Emerges in a Random Walk along Memory [on] [RAM]


Submitted on 1 Feb. 2022 22:45 by
Pavel Tokmakov (TRI)

Running time:0.09 s
Environment:GPU @ 2.5 Ghz (Python)

Method Description:
This paper proposes a self-supervised objective for learning
representations that support localization of objects under occlusion -
a property

known as object permanence. Instead of directly supervising
locations of invisible objects, we show that object permanence can
emerge by

optimizing for a Markov random walk along a space-time graph of
memories, provided that the states in each time step are (non-
Markov)

features from a recurrent encoder.
Parameters:
Will be made available with the manuscript.
Latex Bibtex:
@inproceedings{tokmakov2022object,

title={Object Permanence Emerges in a Random Walk along
Memory},

author={Tokmakov, Pavel and Jabri, Allan and Li, Jie and Gaidon,
Adrien},

booktitle={ICML},

year={2022}

}

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 79.53 % 78.79 % 80.94 % 82.54 % 86.33 % 84.21 % 88.77 % 87.15 %
PEDESTRIAN 52.71 % 53.55 % 52.19 % 58.86 % 70.17 % 59.49 % 64.99 % 77.70 %

Benchmark TP FP FN
CAR 32298 2094 583
PEDESTRIAN 17756 5394 1660

Benchmark MOTA MOTP MODA IDSW sMOTA
CAR 91.61 % 85.79 % 92.22 % 210 78.26 %
PEDESTRIAN 68.40 % 73.61 % 69.53 % 262 48.16 %

Benchmark MT rate PT rate ML rate FRAG
CAR 86.31 % 11.23 % 2.46 % 158
PEDESTRIAN 51.55 % 34.71 % 13.75 % 738

Benchmark # Dets # Tracks
CAR 32881 713
PEDESTRIAN 19416 331

This table as LaTeX