Benchmark for anonymous video analytics

EURASIP Journal on Image and Video Processing

Table 2 Anonymous video analytics algorithms for localization, age, and gender estimation

Alias	Ref.	Name	Person part	Localization		Age	Gender	Implementation		Count performance			Speed
				Detection	Tracking			GPU	CPU	Instantaneous		Cumulative
										People	OTS
A1	[14]	RetinaFace	Face	\(\checkmark\)				\(\checkmark\)	\(\checkmark\)	+	+++	+	+++
A2	[15]	MTCNN	Face	\(\checkmark\)				\(\checkmark\)	\(\checkmark\)	+	+++	+	+++
A3	[16]	DeepSORT	Person		\(\checkmark\)			\(\checkmark\)	\(\checkmark\)	+++	+	+++	++
A4	[17]	TRMOT	Person		\(\checkmark\)			\(\checkmark\)		+++	+	+++	++
A5	[18]	FaceLib	Face	–	–	\(\checkmark\)	\(\checkmark\)	\(\checkmark\)	\(\checkmark\)	–	–	–	+++
A6	[19]	DEX	Face	–		\(\checkmark\)	\(\checkmark\)	\(\checkmark\)	\(\checkmark\)	–	–	–	+++
C1	Commercial-1		Face		\(\checkmark\)	\(\checkmark\)	\(\checkmark\)		\(\checkmark\)	+	+++	+++	++
C2	Commercial-2		Face		\(\checkmark\)	\(\checkmark\)	\(\checkmark\)		\(\checkmark\)	+++	+++	+++	++

Algorithms count performance and speed are assessed, on a scale from 1 to 3 (the higher, the better) and regardless of the system used, based on generic algorithmic concepts (e.g., tracking outperforms detection for cumulative counts) and not on experimental results. OTS: opportunity to see. A4 was not compatible with OpenVINO optimization at submission time, thus we run A4 on CPU without optimization