Table 7 The MAE/RMSE of the oracle evaluation (traditional supervised training within Shanghai Tech Part A) and cross-dataset evaluation (training on NWPU-Crowd and test on Shanghai Tech Part A), and the corresponding performance degradations
From: A survey of deep learning methods for density estimation and crowd counting
Method | Oracle (MAE/RMSE) | Cross-dataset evaluation (MAE/RMSE) | Relative errors rise (%) |
|---|---|---|---|
MCNN [1] | 110.2/173.2 | 151.6/217.7 | \(\uparrow\) 37.6/25.7 |
CSRNet [12] | 68.2/115.0 | 111.0/169.2 | \(\uparrow\) 62.6/53.5 |
C3F-VGG [264] | 71.4/115.7 | 96.5/151.6 | \(\uparrow\) 35.2/31.0 |
CANNet [99] | 62.3/100.0 | 83.5/137.4 | \(\uparrow\) 34.0/37.4 |
SCAR [262] | 66.3/114.1 | 96.6/161.5 | \(\uparrow\) 45.7/71.5 |
SFCN† [97] | 71.5/114.3 | 108.8/185.8 | \(\uparrow\) 52.2/62.6 |