Table 1.

Summary of the participating teams detection and segmentation tasks for the crowd-sourced polyp generalisation challenge.

Team name	Algorithm	Backbone	Nature	Choice basis	Data Aug.	Loss	Opt.	Code	No. of parameters (M)
Task I: polyp detection
AIM_CityU18	FCOS	FPN, ResNeXt -101-DCN	ATSS	Accuracy speed	No	Generalized Focal loss	SGD	[d1]	51.0
HoLLYS_ETRI24	Mask R-CNN	FPN ResNet34	Ensemble	Accuracy++	No	Smooth L1	SGD	[d2]	63.75
JIN_ZJU19	YOLOV5	CSPdarknet SPP	Ensemble	speed++	Yes	BECLogits	SGD	[d3]	140.70
GECE_VISION20	EfficientDet	EfficientNet D0-D3	Ensemble	Accuracy	Yes	Focal loss	Adam	[d4]	30.60
Task II: Polyp segmentation
Aggcmab21	DPN92-FPN	DPN92-FPN	Cascaded	Accuracy++	Yes	BCE	SGD	[s1]	75.91
AIM_CityU18	HRNet + LRM	HRNet	MSFF	Accuracy speed	Yes	BCE, DSC	SGD	[s2]	49.90
HoLLYS_ETRI24	Mask R-CNN	ResNet50	Ensemble	Accuracy+ speed+	Yes	Smooth L1	SGD	[s3]	63.75
MLC_SimulaMet22	DivergentNet	TriUNet	Ensemble	Accuracy++	No	BCE, DSC	Adam	[s4]	180.64
Sruniga23	HarDNet68	HarDNet68	Multiscale	Accuracy+ speed++	No	BCE	Adam	[s5]	17.42

All test was done on NVIDIA V100 GPU provided by the organisers. In total 11 different methods are provided together with the nature of these methods and basis of their choice that the teams considered. All codes for each team are available for reproducibility.

FCOS fully convolutional one-stage object detection, FPN feature pyramid network, ATSS adaptive training sample selection.

YOLO You Only Look Once, SGD Stochastic Gradient ‘escent, [d1]–[d4] hyperlinked GitHub repos.

LRM low-rank module, MSFF multi-scale feature fusion, DPN dual path network, FPN feature pyramid network, BCE binary cross entropy

BCE binary cross entropy, DSC dice similarity coefficient, IoU intersection over union, W weighted, SGD Stochastic gradient descent.