How to train with my own data?

[RoyalLii]

I want to use the Superpoint pre-trained model to train on my own data, do I have to perform steps 1 to 6 in the training step?

[rpautrat]

Hi, no I think you can just use the pre-trained model to export a pseudo ground truth on your data (Step 2), and then directly jump to the training of SuperPoint with this pseudo ground truth (Step 6). If the generated pseudo ground truth does not look good enough, you might have to do one iteration of Step 2 - Step 3 - Step 2 to increase its quality.

[RoyalLii]

Thank you very much for your reply. I will try it.

[RoyalLii]

Can I use the pre-trained model is MagicPoint(synthetic) or MagicPoint(coco) in step2 ?

[rpautrat]

You can use both, but I would recommend the MagicPoint one. Since it is purely specialized in detection, its performance on keypoint detection is slightly better than the joint model SuperPoint.

[RoyalLii]

I use my own dataset for training, put my data directly into the coco2014 train directory？ Running step 2,，Why do I use 235 images to generate only 85 .npz

[rpautrat]

Adapting it to your own images will require a bit more work I think. Please check past issues, because this question has been asked many times.

[RoyalLii]

Thank you for your reply Oh, I'm sorry I didn't find a relevant question，I wanted to ask how long it takes to iterate one epoch, I'm stuck in the second round of training, using a graphics card is 1080ti

[rpautrat]

The epoch duration depends on the size of your dataset, so I don't know how big yours is :)

[foxkw]

Hi, no I think you can just use the pre-trained model to export a pseudo ground truth on your data (Step 2), and then directly jump to the training of SuperPoint with this pseudo ground truth (Step 6). If the generated pseudo ground truth does not look good enough, you might have to do one iteration of Step 2 - Step 3 - Step 2 to increase its quality.

When I do this, I find that the number of labels generated will be reduced compared with that before, for example, on the vehicle/sign board

[rpautrat]

Having a higher number of labels does not necessarily mean that it is of higher quality. But if you prefer having more keypoints, you can always decrease the detection threshold to keep more points.

[foxkw]

Having a higher number of labels does not necessarily mean that it is of higher quality. But if you prefer having more keypoints, you can always decrease the detection threshold to keep more points.

Thanks for your reply. I have a question: Different resolution images have different sizes, some are large, some are small, how to set the topk and threshold parameters to obtain stable feature points, what do you think about this

[rpautrat]

One simple solution would be to keep always the same threshold, and not to enforce any topk. You will get more points on larger images, but since the threshold is the same, you should get similar points. Of course, SuperPoint is not very scale-invariant, so if you have big scale changes, there is no way you can get very stable keypoints in any case.

[foxkw]

One simple solution would be to keep always the same threshold, and not to enforce any topk. You will get more points on larger images, but since the threshold is the same, you should get similar points. Of course, SuperPoint is not very scale-invariant, so if you have big scale changes, there is no way you can get very stable keypoints in any case.

I found that the magipoint model will miss some points when export the label on images, so that the corner detection of the lane is unstable in the continuous frame image. This seems to indicate that magipoint also has shortcomings in the export label stage. Do you have any good ideas, improve it, and look forward to your reply

[rpautrat]

Of course, MagicPoint is not perfect and will always make mistakes. The homography adaptation process is one way to improve the export, but it has its limits. Improving the keypoint ground truth is actually an ongoing area of research, so I cannot help much on this for now :)