strange descriptor loss

[XhqGlorry11]

@rpautrat Congrats to your nice work! I'm trying to reproduce training superpoint-coco but run into strange descriptor loss. I used pretrained magicpoint-coco to export coco labels and train superpoint from scratch.
I set: lambda_d = 800, lambda_loss = 1 and other settings same with superpoint-coco.yaml.
My positive_dist starts around 1.8 and negative_dist starts around 1e-6. But in your training log, positive_dist starts around 1 and negative_dist starts more than 0.2.

Can you help me check what may be wrong? Thx~~
Did you train superpoint-coco from scratch or use any pretrained models?

[rpautrat]

Hi,

I trained superpoint-coco from scratch, so the initial values of positive_dist and negative_dist are not really meaningful: it depends on the randomization of the initialization weights.

What matters are the final values, and I think that your values are quite close to my training. I get on average positive_dist = 0.01 and negative_dist = 0.04.

[XhqGlorry11]

@rpautrat Thank you for your reply. It seems that you set lambda_d=0.05, lambda_loss = 10000 in superpoint_coco.yaml, which is quite different with paper's setting. What's the main reason?

[rpautrat]

The values from the official paper did not work for me. I had to tune these parameters, and the lambda_d = 800, lambda_loss = 1 were the best parameters I found for our released superpoint model.

Since then, I added a normalization in the descriptor loss (95d1cfd) and adapted the values to lambda_d=0.05, lambda_loss = 10000, again by tuning the parameters. So I guess these values should be the right ones to use. Feel free to tune them more of course!

[GuoBo98]

The values from the official paper did not work for me. I had to tune these parameters, and the lambda_d = 800, lambda_loss = 1 were the best parameters I found for our released superpoint model.

Since then, I added a normalization in the descriptor loss (95d1cfd) and adapted the values to lambda_d=0.05, lambda_loss = 10000, again by tuning the parameters. So I guess these values should be the right ones to use. Feel free to tune them more of course!

Hello, may I ask why you added a normalization in the descriptor loss ?

[rpautrat]

There was a short explanation for this here: #95 (comment)