You signed in with another tab or window. Reload to refresh your session.You signed out in another tab or window. Reload to refresh your session.You switched accounts on another tab or window. Reload to refresh your session.Dismiss alert
I used 25 photos for training and set the height and width of the input image to 1080×1920 in the colmap_eval.yaml file. The number of training iterations is 40000 using RTX3080. The study found that the training time takes 50 minutes. Is this a normal training time? Looking forward to your reply, which is very important to me
The text was updated successfully, but these errors were encountered:
Hi, on a single RTX3090, when the image resolution is 512x640 and the batch_size is 1, it takes about 5 minutes per epoch(1000 iteration). Therefore, when the number of training iterations is 40000, it takes about 200 minutes. However, your image resolution is 1080×1920, and the training time is expected to exceed 200 minutes. But from your description and uploaded image, it only takes 50 minutes, which is strange. Besides, our model requires that H and W be integer multiples of 32. Therefore, if the height in the configuration file is set to 1080, an error will be reported. To sum up, I cannot judge whether 50 minutes is a normal time for your case. If you can provide more information, such as where specifically you modified the configuration, I would be willing to help you further.
I used 25 photos for training and set the height and width of the input image to 1080×1920 in the colmap_eval.yaml file. The number of training iterations is 40000 using RTX3080. The study found that the training time takes 50 minutes. Is this a normal training time? Looking forward to your reply, which is very important to me

The text was updated successfully, but these errors were encountered: