IExplainStuff


Project maintained by aflah02 Hosted on GitHub Pages — Theme by mattgraham

Quasi-Dense Similarity Learning for Multiple Object Tracking

Abstract

1. Introduction


Figure 1



How to Tackle Missing Targets -

How to Tackle Multiple Targets -



Benchmarks Attained by the Model -

Location and Motion in MOT

Appearence Similarity in MOT

Nearest Neighbour Search

Constrastive Learning

3. Methodology

RPN and RoIs

Loss Function

3.2 Quasi-dense Similarity Learning


Figure 2


3.3 Object Association

It is not at all trivial how we’re going to use all this stuff we’ve built so far to track objects. Heck, does it even make sense aren’t all we doing just some boring maths?

Let’s try to make some sense of what we are doing


Figure 3

Figure 3

Bi-directional Softmax

No Target Cases

Multi-Target Classes

NMS

NMS

4. Experiments

4.1 Datsets

MOT

BDD100K

Waymo

TAO

Long Tail Distribution

4.2 Implementation Details

ResNet-50

IoU Balanced Sampling

4.3 Main Results

MOT

BDD100K

Waymo

TAO

4.4 Ablation Studies

Importance of Quasi-Dense Matching