TasselNetv2: in-field counting of wheat spikes with context-augmented local regression networks

Table 7 Comparison with state-of-the-art counting approaches on the test set of WSC dataset. TasselNetv2 adopts an AlexNet-like architecture in Fig. 6 and is trained from scratch

Method	Henan Zhengzhou (2012–2013)		Shandong Taian (2012–2013 Camera1)		Overall		#Parameters
Method	MAE	RMSE	MAE	RMSE	MAE	RMSE	#Parameters
Segmentation method in [13]	387.09	436.84	268.03	345.78	317.19	386.22	\(\times\)
CCNN [6]	168.41	214.41	52.40	72.78	101.39	149.91	\(5.70\times 10^5\)
MCNN [5]	149.44	188.34	58.83	75.50	97.08	135.17	\(1.33\times 10^5\)
CSRNet\(^\dagger\) [23]	64.19	88.96	33.26	46.19	46.32	67.63	\(1.63\times 10^7\)
TasselNet [9]	94.97	137.24	36.79	57.37	61.35	99.27	\(6.38\times 10^5\)
TasselNetv2	74.97	113.21	33.12	49.26	50.79	80.66	\(6.38\times 10^5\)
TasselNetv2\(^\dagger\)	61.57	87.67	31.62	47.55	44.27	67.47	\(1.60\times 10^7\)

\(^\dagger\) means the model is finetuned from the pretrained VGG16, and layer-by-layer settings can be found in Additional file. The best performance is italics

ISSN: 1746-4811