Multiple Target Tracking

This page is under reconstruction. I just put the reference which was important to me here. With a bit of description.

Tracking The Untrackable: Learning to Track Multiple Cues with Long-Term Dependencies” of Amir Sadeghian.

There is good blog post about this article.

RNN network, here LSTM, used for each appearance, motion and interaction cues. They are trained separately as well. The combination of all cues are also passed through other RNN and the combined network is trained by fine-tuning the previous network.

The appearance is the main cue. This method uses the previous frames, here 6. This past information improve the result. Detection should be given for appearance network.

Detect to Track and Track to Detect” (1710.03958) (Zisserman)

Advantage: Detection is done on whole image.

Disadvantage: The data association is done by viterbi method.

Re3 : Real-Time Recurrent Regression Networks for Visual Tracking of Generic Objects (1705.06368) (Dieter Fox, Ali Farhadi)

This method is fast, end-to-end, robust to occlusion and appearance changes.

Learning Policies for Adaptive Tracking with Deep Feature Cascades (1708.02973)

By formulating the tracking problem as a decision making process, we learn a reinforcement learning agent that can make such distinctions.

It is an efficient method. Not sure if it is useful!

Learning Hierarchical Features for Visual Object Tracking with Recursive Neural Networks (1801.02021)

It is an efficient method. Not sure if it is useful!

Deep Tracking on the Move: Learning to Track the World from a Moving Vehicle using Recurrent Neural Networks (https://arxiv.org/pdf/1609.09365.pdf), (Ingmar Posner)

This work only on occupancy grid, which is here obtained from laser data. The paper is based on previous work on deep tracking (1602.00991, 1604.05091).

main architecture based on 1604.05091

Online Multi-Target Tracking Using Recurrent Neural Networks (1604.03635) (Konrad Schindler)

This paper doesn’t outperform traditional methods. Detections should be given as input.

Learning by tracking: Siamese CNN for robust target association (1604.07866) (Konrad Schindler)

This is not end to end network.

Benchmark

Tracking the Trackers: An Analysis of the State of the Art in Multiple Object Tracking (https://arxiv.org/pdf/1704.02781.pdf) (Konrad Schindler, Daniel Cremers, Stefan Roth)

Source: Deep Learning on Medium