search

tum-vision / dvo_slam

Dense Visual Odometry and SLAM
632 stars 306 forks source link

Settings to reproduce paper results. #13

Open jekekrantz opened 10 years ago

jekekrantz commented 10 years ago

Hi, Im looking to reproduce the results from your paper "Dense Visual SLAM for RGB-D Cameras". However i seem to be getting very different results than the paper. For example fr1/room i get screenshot from 2014-06-05 12 24 26

where as the paper gives a rmse error of 0.053 m. I assume there is some sort of settings that i do not have the same as in the paper. Could you provide the settings used in the paper?

Cheers, Johan Ekekrantz

christiankerl commented 10 years ago

Hi, you have to run the final dense pose graph optimization. In your plot you can see that the keyframe poses are optimized/corrected, i.e., the spikes in the trajectory. To also move the intermediary frames to their correct poses you can run a batch graph optimization at the end. Open dynamic_reconfigure and go to dvo_vis/graph. there is a checkbox final_graph_opt. make sure to start dvo_benchmark with the keep_alive:=true option, otherwise it will terminate without this final optimization.