ALFPN: Adaptive Learning Feature Pyramid Network for Small Object Detection

Table 5

Comparison with the state-of-the-art methods on TT-100K dataset.

No.	Method	Dataset	Module	AP
1	Faster R-CNN	TT-100K	FPN [9]	43.3	27.2	48.1	56.1
2			PAFPN [48]	40.6	22.7	44	54.6
3			HRFPN [49]	43.4	27.6	47.5	56.3
4			AugFPN [26]	45.6	29.3	49.2	58.4
5	Cascade R-CNN	TT-100K	FPN [9]	44.7	28.3	48.4	58.6
6			PAFPN [48]	43.3	25.3	45.1	55.2
7			HRFPN [49]	45.8	26.5	47.3	54.5
8			AugFPN [26]	46.3	30.1	48.4	59.6
9	Sparse R-CNN	TT-100K	FPN [9]	57.3	38.5	62.1	69.8
10			PAFPN [48]	57.9	38.1	65.2	68.6
11			HRFPN [49]	55.6	37.3	60.4	63.1
12			AugFPN [26]	58.7	39.8	61.9	70.3
13	Retinanet	TT-100K	FPN [9]	34.8	22.1	39.2	43.1
14			PAFPN [48]	35.2	23.2	39.1	43.8
15			HRFPN [49]	34.3	22.3	38.1	42.9
16			AugFPN [26]	36.1	23.4	39.5	43.2
18	Retinanet (ours)	TT-100K	ALFPN	37.4	25.1	40.3	44.9
19	Cascade R-CNN (ours)		ALFPN	48.6	31.9	49.8	62.5
20	Faster R-CNN (ours)		ALFPN	48.7	32.3	50.4	61.4
21	Sparse R-CNN (ours)		ALFPN	60.5	42.6	63.0	72.6

The symbol means the experimental results of our proposed module.