Benchmark:

MOT17 |

Short name:

RGCN_T

Detector:

Public

Description:

we present a unified graph optimization framework to solve the data as- sociation problem in multiple object tracking. The proposed framework build graphs from detection nodes and formulate the graph optimization to a learnable link prediction problem. Specifically, we first use a relational graph convolution net- work (R-GCN) to learn features of both detection nodes and their linked edges. To exploit various edge information, we then use a message passing network (MPN) to aggregate dif- ferent types of edge representations and propagate them on the graph iteratively. Finally, we build an end-to-end learn- able link prediction model by using the leaned edge features to predict if two nodes should be linked. Experimental re- sults on three benchmark datasets demonstrate the effective- ness of the proposed framework, outperforming state-of-the- art methods in multi-object tracking.

Reference:

Last submitted:

October 02, 2020 (1 year ago)

Published:

October 12, 2020 at 08:40:10 CET

Submissions:

3

Project page / code:

n/a

Open source:

No

Hardware:

3 GHZ

Runtime:

59.2 Hz

Benchmark performance:

Detailed performance:

Raw data:

download