training problem

[chaoxianluo]

Hello, I trained this network using scene fire in 7Scenes dataset. I firstly trained for 10 epochs, but the result is not very good. Then I trained another 40 epochs(50 in total) and tested it using the test images. Results show that the network outputs same pose for every picture.I watched the log in tensorboard and found all parameter sx and sq kept dropping. I wonder if this is a normal state or there is someting wrong. I noticed in paper posenet2, initial sx and sq are 0.0 and -3.0. Howerver in your code they are both -3.0. I don't know what error I made during training. I would be grateful if you have any suggestion on this problem, or you can provide more training detail.Thank you very much

[JBLanier]

Hi there! It's a year later, but I'm experiencing the exact same problem.

Did you ever get this working (and remember what you did)?

Thanks.

[chenusc11]

@chaoxianluo I tried to reimplement the VLocNet myself for the past few weeks and I found the same phenomenon with sx and sq kept dropping. After some experiment and comparison with the official MapNet code, I think this is normal. The only weird thing is the geometric consistency loss gives a negative value, but I believe this is the correct implementation.

In terms of the initial sx and sq initialization, I guess this author just uploaded what he was using. However, I found even if you chose 0.0 and -3.0 as in the PoseNet paper, the result won't be anything better.

Personally, I think the author did a fairly good job re-implementing the paper code that I have found very little place which wasn't the same as the paper (ie. the previous pose fusion took place at 7x7 feature map size, but 14x14 in this repo). After all, I think the main performance problem should be somewhere else. Maybe some key elements we missed from the paper