Object Tracking Models

List of Object Tracking Models

The table below shows the object tracking models available for each task category.

Category	Model	Documentation
General	IoU Tracker	dabble.tracking
	OpenCV MOSSE Tracker	dabble.tracking
Human	JDE	model.jde
	FairMOT	model.fairmot

Benchmarks

Inference Speed

The table below shows the frames per second (FPS) of each model type.

Model	Object Detector Type	Input Size	CPU	GPU
IoU Tracker with YOLOX	yolox-m	–	7.87	36.18
OpenCV MOSSE Tracker with YOLOX	yolox-m	–	6.74	21.45
JDE	–	–	1.86	26.32
FairMOT	–	864 × 480	0.30	22.60

Hardware

The following hardware were used to conduct the FPS benchmarks:

- CPU: 2.8 GHz 4-Core Intel Xeon (Cascade Lake) CPU and 16GB RAM

- GPU: NVIDIA A100, paired with 2.2 GHz 6-Core Intel Xeon CPU and 85GB RAM

Test Conditions

The following test conditions were followed:

- input.visual, the model of interest, and dabble.fps nodes were used to perform inference on videos

- A video sequence from the MOT Challenge dataset (MOT16-04) was used

- The video sequence has 1050 frames and is encoded at 30 FPS, which translates to about 35 seconds

- 1280×720 (HD ready) resolution was used, as a bridge between 640×480 (VGA) of poorer quality webcams, and 1920×1080 (Full HD) of CCTVs

Model Accuracy

The table below shows the performance of our object tracking models using multiple object tracker (MOT) metrics from MOT Challenge. Description of these metrics can be found here.

Model	Object Detector Type	MOTA	IDF1	ID Sw.	FP	FN
IoU Tracker with YOLOX	yolox-m	34.1	40.9	960	8997	62830
OpenCV MOSSE Tracker with YOLOX	yolox-m	32.8	38	2349	7695	65268
JDE	–	70.1	65.1	1321	6412	25292
FairMOT	–	81.8	80.9	536	3663	15903

Dataset

The MOT16 (train) dataset is used. We integrated the MOT Challenge API into the PeekingDuck pipeline for loading the annotations and evaluating the outputs from the models. MOTA and IDF1 are reported in percentages while IDS, FP, and FN are raw numbers.

Only the “pedestrian” category in MOT16 (train) was processed.