Images found but not added to dataset

[felipesce]

Greetings,

I'm having a very strange error, and reckon I probably have to adjust some threshold but haven't had much luck finding what and where to change.

I'm training with 132 sets of 100 images/depths/poses, for a total of close to 13.000 sets of 7 frames.

While building the dataset with the function _build_dataset, it effectively finds the 100 images, depths and poses for each folder, but at the end of the dataset creation, the function dataset_factory in factory.py states that only 10-250 images were found.

By the definition of __len__(db), in base.py, this means that _build_dataset_index is not adding every image to the frame graph, which is confusing since none of these scenes are designated as test scenes.

Any ideas on how to debug this issue?

Some ideas I've had:

• Co-visibility between close frames is less than 50% (where can I change this threshold?)
• Scenes are too difficult (or easy), however changing the value of fmin & fmax hasn't made much difference.
• Changing max_flow in build_frame_graph. I have no idea what this does, or how to know the correct value.
• Changing MAX_FLOW = 100.0 or s = 2048 or val.mean(-1) < 0.7 in compute_distance_matrix_flow in rgbd_utils.py. Same as above.
• I also tried inspecting how the frame graph is built and what means what, but my confusion increased exponentially.

Any and all comments or insights would be deeply appreciated. Have a nice day.

[felipesce]

Diminishing MAX_FLOW seems to have worked, the images are recognized; I'm not sure if it will have any other effect yet.

[felipesce]

My problem seems to be that the scenes have many parts that are very similar, so varying the threshold either makes very few images pass or too many false positives (images with no co-visibility are believed to be co-visible, leading to poor training.)

I'm at odds as to how to tweak the parameters to diminish this problem, so I'll try to construct a hand-made co-visibility graph to test if this is indeed the issue.

Thanks again to anyone who wants to chime in.