Thank you for your work, I would like to ask if you have pre-trained model parameters here, and if so, can you make them public? I will be very grateful.

Hello,

Thanks for your work.

Can you explicit what are the different settings:

• --model
• --pretrained_model
• --test_model

Thanks.

Great implementation of PoseNet! However, the loss doesn't seem to converge to minima when I start learning sx and sq. Any help with recreating the results in the Geometric Pose Loss paper would be a big help!

Hi,

As the implementation of the loss
loss = torch.exp(-self.sx)*loss_x + self.sx + torch.exp(-self.sq)*loss_q + self.sq
is equivalent to eq.4 in Kendall et al. paper, I still not understand how can the trainable parameters 'sx' and 'sq' can capture the variance of position and orientation noise from the model output? I mean, how come you implement the loss this way?

Thank you very much for your time. :)

Hi, I have a question!

Does sx mean the standard deviation(variance) of error : x-x' ?

I mean.. if we set batch size to 30,

then Sx = stard deviation of 30 x-x'? (Because x has three values (x, y, z), Sx = std of 90 numbers?)

20th train Loss: total loss -1.273 / pos loss 1.371 / ori loss 0.018
21th train Loss: total loss -1.584 / pos loss 1.084 / ori loss 0.016
22th train Loss: total loss -1.778 / pos loss 0.875 / ori loss 0.017
23th train Loss: total loss -1.884 / pos loss 0.764 / ori loss 0.018
24th train Loss: total loss -1.651 / pos loss 1.008 / ori loss 0.017
25th train Loss: total loss -1.831 / pos loss 0.801 / ori loss 0.018
26th train Loss: total loss -1.562 / pos loss 1.067 / ori loss 0.018
27th train Loss: total loss -1.713 / pos loss 0.965 / ori loss 0.016
28th train Loss: total loss -1.338 / pos loss 1.275 / ori loss 0.019
29th train Loss: total loss -1.722 / pos loss 0.964 / ori loss 0.016
30th train Loss: total loss -1.551 / pos loss 1.094 / ori loss 0.018
31th train Loss: total loss -1.603 / pos loss 1.084 / ori loss 0.016
32th train Loss: total loss -1.714 / pos loss 0.947 / ori loss 0.017
33th train Loss: total loss -1.465 / pos loss 1.029 / ori loss 0.025
34th train Loss: total loss -1.740 / pos loss 0.927 / ori loss 0.017
35th train Loss: total loss -1.579 / pos loss 0.987 / ori loss 0.022
36th train Loss: total loss -1.537 / pos loss 1.074 / ori loss 0.019
37th train Loss: total loss -1.515 / pos loss 1.123 / ori loss 0.018
38th train Loss: total loss -1.181 / pos loss 1.409 / ori loss 0.020
0th val Loss: total loss -2.384 / pos loss 0.434 / ori loss 0.009
1th val Loss: total loss -2.452 / pos loss 0.358 / ori loss 0.009
2th val Loss: total loss -2.407 / pos loss 0.375 / ori loss 0.011
3th val Loss: total loss -2.356 / pos loss 0.417 / ori loss 0.011
Train and Validaion error -1.696101188659668 / -2.395916223526001

========================================
Epoch 225/399

0th train Loss: total loss -1.572 / pos loss 1.093 / ori loss 0.017
1th train Loss: total loss -1.546 / pos loss 1.108 / ori loss 0.017
2th train Loss: total loss -1.847 / pos loss 0.873 / ori loss 0.014
3th train Loss: total loss -1.953 / pos loss 0.679 / ori loss 0.018
4th train Loss: total loss -1.753 / pos loss 0.908 / ori loss 0.017
5th train Loss: total loss -1.600 / pos loss 1.083 / ori loss 0.016
6th train Loss: total loss -1.617 / pos loss 1.014 / ori loss 0.018
7th train Loss: total loss -1.760 / pos loss 0.888 / ori loss 0.018
8th train Loss: total loss -1.264 / pos loss 1.346 / ori loss 0.019
9th train Loss: total loss -1.606 / pos loss 1.040 / ori loss 0.018
10th train Loss: total loss -1.619 / pos loss 0.982 / ori loss 0.020
11th train Loss: total loss -1.708 / pos loss 0.935 / ori loss 0.018
12th train Loss: total loss -1.645 / pos loss 1.058 / ori loss 0.015
13th train Loss: total loss -1.703 / pos loss 0.959 / ori loss 0.017
14th train Loss: total loss -1.768 / pos loss 0.860 / ori loss 0.019
15th train Loss: total loss -1.705 / pos loss 0.931 / ori loss 0.018
16th train Loss: total loss -1.712 / pos loss 0.958 / ori loss 0.016
17th train Loss: total loss -1.656 / pos loss 0.974 / ori loss 0.018
18th train Loss: total loss -1.686 / pos loss 0.984 / ori loss 0.016
19th train Loss: total loss -1.657 / pos loss 0.989 / ori loss 0.018

pos out [ 60.13612 -37.06963 1.4968748]
ori_out [ 1.43032163 0.86646476 -0.1398881 ]
pos true [ 58.309566 -36.34202 1.250287]
ori true [ 1.38532974 0.91265223 -0.18131388]
[ 60.13612 -37.06963 1.4968748]
[ 58.309566 -36.34202 1.250287]
1219th Error: pos error 1.982 / ori error 0.077

Overall median pose errer 0.973 / 0.054
Overall average pose errer 1.162 / 0.069

Hi,
Great work on the repo!
I couldn't find any links/reference to any pretrained model files for a quick test.

Thanks

Do you implement geometric loss function?

In the bayesian posenet article "Modelling Uncertainty in Deep Learning for Camera Relocalization", it is given that dropout layers are added after each convolutional layer, not at the last layer. Please fix that.

I tried both the wget command, and going to the link that directed me to the University of Cambridge site, but I got a proxy error for both attempts.

Is there any way to download the dataset?

Hi!!! Thank you for your contribution.
The freshmen have some problems.
What does the test generated file mean?
What do each column of parameters mean? 6 parameters.
Looking forward to your reply.