Loping151 / EndoGSLAM

Offical Code Implementation for EndoGSLAM
https://loping151.github.io/endogslam
41 stars 0 forks source link

Question regarding the keyframes and the length of gt_train_w2c #9

Open bsticker opened 1 month ago

bsticker commented 1 month ago

It looks like the poses of the keyfiles are not saved. If I change the keyframe_every to a different number, it does not give any other result either. Do you have a suggestion how I can see these keyfiles? Also I do not understand why the gt_train_w2c is not the same length as the groundtruth file (pose.txt) but shorter. It is the same length as est_w2c.txt. How do I change the length of the file and what makes the est_w2c.txt shorter than the number of depth/color images it was given? In the eval_save function the num_frames is defined by the length of the dataset, but this makes me think that it should be as long as the total amount of color/depth images there are (and thus the length of the ground truth pose.txt file).

Loping151 commented 2 weeks ago

I'm so sorry for the very late reply. Somehow I failed to receive emails for this repo.

What do you mean keyfiles? For the poses, all the poses for train indices should be saved in est_w2c.

About the length: there is a train-test split used when evaluating the nvs quality: utils/eval_helpers.py

            if (total_time_idx + 1) % 8 != 0:
                dataset_type = 'train'
                _dataset = train_dataset
                time_idx = total_time_idx - (total_time_idx // 8)
            else:
                dataset_type = 'test'
                _dataset = eval_dataset
                time_idx = total_time_idx//8

7 of every 8 frames are feed into endogslam, and the gt_train_w2c is the corresponding poses on the training indices for evaluation. It is only for evaluating the traj, ensuring the indices are the same as est_w2c.