Table 5.

Experiments of discussion on shape and texture joint learning.

Backbone layers	Structure	Acc.↑	Pre.↑	Rec.↑	F1↑
50	Single-streama	0.950	0.872	0.945	0.907
	Cascade Cls. and Seg.b	0.950	0.886	0.928	0.907
	Two-stream without joint learningc	0.960	0.909	0.939	0.924
	Two-stream with joint learningd	0.967	0.904	0.977	0.939
101	Single-streama	0.952	0.877	0.950	0.912
	Cascade Cls. and Seg.b	0.955	0.888	0.945	0.916
	Two-stream without joint learningc	0.961	0.911	0.944	0.927
	Two-stream with joint learningd	0.971	0.916	0.978	0.946

^aSingle-stream: only use the texture encoder in the proposed method for feature extraction.

^bCascade Cls. and Seg.: cascading segmentation network in front of classification network.

^cTwo-stream without joint learning: removing the feature decoder in the shape-biased stream of our method.

^dTwo-stream with joint learning: the proposed framework.