Table 2.

Ablation experiments of surgical gesture recognition on suturing task averaged over eight cross-validation runs under LOUO scheme. Acc., Edit and F1@{10, 25, 50}, represent the frame-wise accuracy, edit distance and F1 score in different thresholds

Suturing (LOUO)	Acc.	Edit	F1@10	F1@25	F1@50
Self-attn only	87.8	44.0	54.8	53.5	49.0
Dilation only	90.1	76.8	81.9	81.5	78.5
Encoder dilation+ attn	90.8	76.9	82.5	81.8	79.3
Decoder dilation +attn	90.5	77.9	83.4	83.4	79.7
Symmetric dilation + attn	90.7	83.7	87.7	86.9	83.6
Symmetric dilation + attn + pooling	90.1	89.9	92.5	92.0	88.2
SD-Net (w. multi-task)	90.5	90.6	93.5	92.8	90.5

Bold values indicate the comparing with all the listed methods, which one reaches the highest evaluation score of that column (evaluation metric)