Table 2.

Comparison with previously published methods based on SSD, WCE, CVC-ClinicDB, and ETIS-Larib test sets. Pre-train means that a pre-trained backbone was adopted to initialize the model or it was initialized from scratch. The speed (FPS) and the (mAP) performances were tested using google Colab pro+ GPU.

Training Data	Methods	Backbone	Input Size	Pre-Train	FPS	mAP@0.5(%)
						WCE
	SSD300	VGG16	300 × 300 × 3	✓	46	77.2
	SSD300	ResNet-101	300 × 300 × 3	✓	47.3	81.65
	SSD500	VGG16	300 × 300 × 3	✓	19	79.45
	SSD500	ResNet-101	300 × 300 × 3	✓	20	84.95
WCE	FSSD300	VGG16	300 × 300 × 3	✓	65.9	89.78
	FSSD500	VGG16	500 × 500 × 3	✓	69.6	88.71
	DF-SSD300 [41]	DenseNet-S-32-1	300 × 300 × 3	✓	11.6	91.24
	L_SSD [58]	ResNet-101	224 × 224 × 3	✓	40	89.98
	MP-FSSD [18]	VGG16	300 × 300 × 3	✓	62.57	93.4
	Hyb-SSDNet (ours)	Inception v4	299 × 299 × 3	✓	44.5	93.29
						CVC-ClinicDB	ETIS-Larib
	SSD300	VGG16	300 × 300 × 3	✓	46	74.5	74.12
	SSD300	ResNet-101	300 × 300 × 3	✓	47.3	78.85	75.73
CVC-ClinicDB	SSD500	VGG16	500 × 500 × 3	✓	19	78.38	75.45
&	SSD500	ResNet-101	500 × 500 × 3	✓	20	82.74	80.14
ETIS-Larib	FSSD300	VGG16	300 × 300 × 3	✓	65.9	87.26	86.3
	FSSD500	VGG16	500 × 500 × 3	✓	69.6	87.54	86.92
	DF-SSD300 [41]	DenseNet-S-32-1	300 × 300 × 3	✓	11.6	89.92	86.84
	L_SSD [58]	ResNet-101	224 × 224 × 3	✓	40	88.18	87.23
	MP-FSSD [18]	VGG16	300 × 300 × 3	✓	62.57	89.82	90
	Hyb-SSDNet (ours)	Inception v4	299 × 299 × 3	✓	44.5	91.93	91.10