Skip to main content

Table 2 Influence of the proposed architectural decisions on the detection network efficiency

From: Traffic flow estimation with data from a video surveillance camera

 

Baseline

 

Overall gain

Mask branch

✗

✓

✓

✓

✓

 

Anchors optimization

✗

✗

✓

✓

✓

 

Focal loss

✗

✗

✗

✓

✓

 

Adaptive feature pooling

✗

✗

✗

✗

✓

 

\(AP^{bb}\)

57.3

58.2

58.4

58.9

59.2

+1.9

\(AP^{bb}_{50}\)

79.0

79.7

79.7

79.7

80.3

+1.3

\(AP^{bb}_{75}\)

67.5

68.8

68.9

69.4

70.0

+2.5

\(AP^{segm}\)

–

56.1

56.4

56.7

57.2

+1.1

\(AP^{segm}_{50}\)

–

78.9

79.4

79.5

80.2

+1.3

\(AP^{segm}_{75}\)

–

68.0

69.2

68.8

69.1

+1.1

  1. Italic values indicate the performance gain relative to the Faster R-CNN baseline when applying all the proposed improvements