gazeto3d

[liushifu12138]

(pytorch3d) C:\Users\Admin\Desktop\GAZE\PnP-GA-main>python test.py --i -1 --p 0 --savepath path/to/save --target mpii
Test Set: cross-mpii
Read data
[Read Data]: Total num: 37652
Model building
Test 1
Traceback (most recent call last):
File "C:\Users\Admin\Desktop\GAZE\PnP-GA-main\test.py", line 118, in
accs += angular(gaze, gts.cpu().numpy()[k])
File "C:\Users\Admin\Desktop\GAZE\PnP-GA-main\test.py", line 31, in angular
label = gazeto3d(label)
File "C:\Users\Admin\Desktop\GAZE\PnP-GA-main\test.py", line 18, in gazeto3d
gaze_gt[0] = -np.cos(gaze[1]) * np.sin(gaze[0])
IndexError: index 1 is out of bounds for axis 0 with size 1

[MickeyLLG]

It seems that the size of gaze predictions is wrong, could you please check the above tensor "gaze" while testing?

[liushifu12138]

gazes= tensor([[ 8.2086e-02, -2.9102e-01],
[ 1.2727e-01, -2.7274e-01],
[-1.4144e-02, -3.3060e-01],
[ 4.7616e-02, -2.5660e-01],
[-4.9153e-02, -2.9897e-01],
[ 1.1854e-01, -2.8301e-01],
[ 9.7247e-03, -2.9910e-01],
[-2.9545e-02, -2.3393e-01],
[ 3.5883e-02, -2.8869e-01],
[ 9.9566e-02, -1.8370e-01],
[ 2.4126e-02, -3.1156e-01],
[ 6.8830e-02, -3.0675e-01],
[ 3.2630e-02, -2.2405e-01],
[ 9.8224e-02, -2.0690e-01],
[ 9.9875e-02, -2.9076e-01],
[ 2.2173e-02, -1.7108e-01],
[-1.1407e-01, -2.3857e-01],
[ 3.2969e-02, -2.4871e-01],
[ 4.4341e-03, -2.1661e-01],
[-3.9321e-03, -1.7889e-01],
[-4.0874e-02, -2.0264e-01],
[-1.3500e-02, -1.7941e-01],
[ 5.0466e-02, -1.4337e-01],
[ 3.6743e-02, -1.5934e-01],
[ 4.9355e-02, -1.7999e-01],
[ 1.6714e-01, -1.5121e-01],
[ 1.0406e-01, -1.6036e-01],
[ 5.1303e-02, -1.4854e-01],
[ 7.3057e-02, -3.0262e-01],
[-4.4771e-02, -2.6593e-01],
[-3.4157e-02, -2.8508e-01],
[-1.2657e-02, -2.0818e-01],
[ 9.1158e-02, -1.4793e-01],
[ 8.7987e-02, -2.5586e-01],
[ 1.0863e-02, -2.8685e-01],
[ 5.1118e-02, -2.7291e-01],
[ 1.3355e-01, -1.9285e-01],
[ 9.0142e-02, -2.1928e-01],
[ 1.1850e-01, -2.1832e-01],
[ 1.7620e-01, -3.6881e-01],
[ 9.8562e-02, -2.3987e-01],
[ 6.8570e-02, -2.4656e-01],
[ 9.8616e-02, -2.2655e-01],
[ 1.5842e-01, -2.3868e-01],
[-1.8541e-02, -2.4785e-01],
[ 8.2156e-02, -2.3449e-01],
[ 1.1603e-01, -2.7651e-01],
[ 6.5590e-03, -2.9033e-01],
[ 9.7916e-02, -1.7554e-01],
[ 2.7684e-02, -1.8274e-01],
[ 2.4268e-02, -1.8580e-01],
[-4.5271e-02, -1.6706e-01],
[ 2.9319e-02, -1.6746e-01],
[ 4.6420e-02, -1.3384e-01],
[ 6.4820e-02, -1.5720e-01],
[ 3.3350e-02, -1.3427e-01],
[ 3.0542e-02, -1.7560e-01],
[ 5.1842e-02, -1.6244e-01],
[ 7.5468e-02, -1.3210e-01],
[ 8.0944e-02, -1.1087e-01],
[-1.1317e-02, -2.3868e-01],
[ 5.9044e-02, -1.8055e-01],
[ 1.2062e-01, -2.3585e-01],
[ 8.7710e-02, -2.4554e-01],
[ 2.6817e-03, -1.7670e-01],
[ 2.8156e-02, -2.2218e-01],
[ 7.2176e-02, -1.4259e-01],
[ 4.2038e-02, -1.4522e-01],
[ 6.4642e-02, -1.9856e-01],
[ 5.2133e-02, -1.3357e-01]], device='cuda:0')

[liushifu12138]

gts= tensor([[ 126.],
[ 211.],
[ 878.],
[1092.],
[ 441.],
[1222.],
[ 421.],
[ 248.],
[1120.],
[ 605.],
[ 194.],
[1222.],
[ 135.],
[1272.],
[ 379.],
[ 471.],
[ 713.],
[1225.],
[ 989.],
[ 671.],
[1159.],
[ 637.],
[ 735.],
[1013.],
[ 419.],
[ 202.],
[ 365.],
[ 446.],
[1004.],
[ 701.],
[ 685.],
[1073.],
[1155.],
[1133.],
[ 920.],
[ 856.],
[ 143.],
[1013.],
[ 815.],
[ 708.],
[ 332.],
[ 963.],
[ 221.],
[ 124.],
[1213.],
[1276.],
[ 176.],
[ 479.],
[ 709.],
[ 667.],
[ 212.],
[ 456.],
[1178.],
[ 564.],
[ 156.],
[ 601.],
[ 289.],
[ 139.],
[ 541.],
[1275.],
[1079.],
[ 903.],
[ 587.],
[1216.],
[1180.],
[1050.],
[ 694.],
[ 970.],
[ 529.],
[ 329.],
[ 377.],
[ 437.],
[ 180.],
[ 71.],
[ 642.],
[ 799.],
[ 417.],
[ 575.],
[ 742.],
[ 162.]], device='cuda:0')
WHY？？？？

[MickeyLLG]

It is correct the size of "gazes" is like what you've shown, but the labels are obviously wrong. I think you might read the wrong part of your mpii gaze label, probably editing your mpii label file or the reader.py, and making sure they suit well would help.

[liushifu12138]

how to get picture of gaze estimation in test?

[MickeyLLG]

If you have any problems in preparing a dataset, please contact the authors of some common papers such as ETH, MPII, and Gaze360. By the way, the pre-processing method (normalization) is illustrated in "MPIIGaze: Real-World Dataset and Deep Appearance-Based Gaze Estimation". If you have no idea how to read images from your disks, please refer to the reader.py#L74, you can also use cv2.imshow during testing for visulization.

[liushifu12138]

thanks