Why the result of KITTI is worse than presented in the paper

[EthanYee]

Hi, I ran orbslam in catkin workspace and tested some sequences of KITTI. However, most of the results were worse than those on the paper. By the way, I have modified "Settings.yaml" on camera parameters. Does anybody came across this problem ever? A reference camera path result compared with ground truth is: http://pan.baidu.com/s/1hrkTDAG (download link, sorry I failed to upload image here)

[amiltonwong]

What configurations does your machine have? The baseline is i7 cpu

[EthanYee]

My laptop runs on i5 cpu. But I have added some code to make the program working with KITTI image sequences, which means each frame will be processed. Does the computer configuration matter in this condition?

[amiltonwong]

See: https://github.com/raulmur/ORB_SLAM/issues/2#issuecomment-83440067

[EthanYee]

@amiltonwong , thanks very much for you reply. I read the #2 comment but I think my problem is different. They came across with "LOST TRACKING" failure because of weak cpu ( i3 ) and suggested rosbag test. However, I tested the image sequence frame by frame without transferring the images to .rosbag file. No matter how long it takes to tracking a new frame, the consecutive frame won't be drop out.

I didn't change the other two threads( LocalMapping and LoogClosing ), does it matters?

ps: I used the following function for image sequence test which will be called in Tracking::Run();

void Tracking::simuWithSeqImg(std::string imgdir, std::string timefile) { double curtime; char name[200]; int count = 0; std::ifstream timein(timefile.c_str()); ros::Rate r(5); while( ros::ok() ) { double time; std::sprintf(name, "%06d.png", count++); std::string imgname = imgdir + "/" + std::string(name); cv::Mat im = cv::imread(imgname, cv::IMREAD_GRAYSCALE); if (im.empty() ) break; timein>>time; // process same as function "GrabImage()" r.sleep(); } }