Open mkocabas opened 2 years ago
The average confidence weights across the whole trajectory might not be the best indicator of camera pose estimation quality, since a single incorrect prediction in the middle of the video is enough for the rest of the trajectory to veer off.
Have you tried measuring the minimum average-weight over all optimization windows? This might indicate if a mistake was made at any point. Also, the solver residuals may be helpful as well.
Hi @mkocabas Can you tell me where is the screen shot picture you put from? I am also interested in pose estimation confidence.Thanks
Hi @lahavlipson @zachteed!
Great work, thanks a lot for releasing such a neat implementation!
I wonder whether it is possible to obtain confidences in a way that is correlated with the camera pose estimation quality. In an ideal scenario, I need the per-frame confidences which captures the uncertainty of camera localization to design an objective. I noticed that there are confidence weights estimated for each edge in the factor graph. But they were not quite helpful when I compare them across frames in a video. Also between two videos (one with bad, another with good camera localization results), the average confidence weights of the video with bad camera localization was higher. Do you have any idea how I can capture such uncertainty from the model?