Self-Recurrent Learning and Gap Sample Feature Synthesis-Based Object Detection Method

Table 3

Fusion results on the COCO2017 dataset.

Model	Schedule	Image size	Recall	Precision	[email protected]		GIoU	Obj	cls
FPN [14]	×1	416	50.82	38.40	53.40	41.58	0.045	0.793	0.532
FPN-3	×3	416	48.76	32.58	46.58	37.08	0.115	2.232	0.240
AsFF [18]	×1	416	81.56	48.73	75.76	59.00	0.041	1.349	0.398
Passp [41]	×1	416	62.31	38.13	59.13	44.73	0.138	1.933	0.305
Pacsp [41]	×1	416	56.13	40.58	56.38	44.67	0.065	1.727	0.308
NAS-FPN [19]	×1	416	34.03	16.86	28.10	19.47	0.498	4.750	0.110
AugFPN [42]	×1	416	72.37	35.58	67.60	45.57	0.025	2.14	0.300
BiFPN-1 [20]	×1	416	26.01	19.93	28.43	20.9	0.567	8.360	0.315
BiFPN-3	×3	416	68.05	46.57	70.73	53.45	0.202	1.357	0.464
Ours	×1	416	77.95	61.97	79.07	65.00	0.053	1.20	0.331

We compared object detection results on current popular fusion modules to demonstrate the effectiveness of our model. Ours: SLFF + REAML + GSFF.