hajimetaira / inloc_demo Goto Github PK

How can I calculate the angular error like the table in the paper?
for example, the true 6 DoF is X, Y, Z, angle_x, angle_y, angle_z
the estimated 6 DoF is X1, Y1, Z1, angle_x1, angle_y1, angle_z1.

Hi I am playing with Inloc database image poses and I have a question about pose calculated by P_db.m (please also refer to this discussion).

I found there is a difference (about 0-0.5m in translation and 0-1 deg in rotation) between pose calculated by P_db.m and pose calculated by PnP method (pycolmap) using all 2D points from database image and the corresponding aligned 3D points. I build 3D points by referring hloc and I used the camera intrinsic of "SIMPLE_PINHOLE, 1600, 1200, 1385.6, 800, 600" which I believe is correct.
Also, I use aligned 3D points and two poses (one from P_db.m, another from PnP) to render images. I found the rendered image from PnP is more close to original image while rendered image from P_db.m has an obvious shift (I can share the images if necessary).

Do you have any insights of why this happened? I feel translation difference about 0.5m is huge, especially under current Inloc localization threshold: (0.25m, 10°) / (0.5m, 10°) / (1m, 10°). I think pose from P_db.m totally makes sense, but I didn't see anything wrong with pose calculated from PnP.

Thanks!

Hi, @HajimeTaira

1. Is the synthsis process time-costly? Have you ever evaluated the synthesis time?

2. I'm curious about the total time that TrainPV needs.

Really hope to get your reply. Thanks in advance and have a nice day :)

InLoc_demo assumes score array at the input. The score array determines the similarity between query-cutout pairs. The scores are presumably computed by first computing features of the query/cutout images and computing their dot product.

I have provided an implementation of how the input scores are built. However, it would be nice to include it also in this repository for reference. If you agree with my implementation, feel free to copy any of my code and use it in this repository.

Best regards,
Pavel

Thanks for your work and for making the source code publicly available.
I have noticed that the dense feature matching step takes close to 2 seconds per image pair, i.e. the matching against 100 images takes longer than three minutes for one query image. Is there anything we can do to speed it up?

Hello,
If I only want to test the performance of feature detectors and descriptors without using depth information, can it be achieved?

Hello,

is there a possibility to upload the ground truth data?
Also, is there any plan on releasing the evaluation tool ?

Thank you !