Table 3.

Comparison with other methods for pre-segmented Graffiti database

Paper	Feature-type	Features	Classifier	Accuracy
[28]	Hand-crafted	Longest common subsequence (LCS)	HMM, CRF, Most probable LCS (MPLCS)	89.50%, 96.40%, 98.30%
[29]	Hand-crafted	Trajectory matching	Max cosine similarity, fastNN	97.60%
[30]	Both hand-crafted and deep features	CRF-based temporal features	CNN and CRF combined	98.40%
Our method	Deep features	Deep network, motion template	Only 2D-CNN, only 3D-CNN, Late fusion	92.60%, 97.30%, 99.20%