search

shichaoy / semantic_3d_mapping

Semantic 3D Occupancy Mapping through Efficient High Order CRFs
Other
194 stars 60 forks source link

Weird result gained. #10

Open YiheChen1995 opened 5 years ago

YiheChen1995 commented 5 years ago

screenshot from 2019-02-12 15-45-50 HI I wonder if you have any idea where the things could have gone wrong, as you can see that things here are not parallel....

shichaoy commented 5 years ago

Hi, this is mostly due to wrong calibration parameters. Rightnow, I hardcode the calib parameters in main_stereo_imgs.cpp. Since you are trying a new dataset, you should change them, either in the main cpp. Or you can modify the code a bit to put parameter in some yaml file, then you don't need to compile it everytime.

qq664956261 commented 5 years ago

Can you send me the author's dataset? The download link failed. Thank you. @YiheChen1995

speed8928 commented 5 years ago

I got a similar weird result as well. I used the raw kitti dataset sequence 2. I changed the code from main_stereo_imgs.cpp and rebuild it.

image_width = 1242; image_height = 375; calibration_mat<<984.2439, 0, 690, 0, 980.8141, 233.1966, 0, 0, 1; Screenshot from 2019-04-11 22-06-05

I noticed that the depth map you upload is like an inverse version(the further the brighter). Did you do post-processing on it?

speed8928 commented 5 years ago

Hi, I just figure it out the problem is the depth map should be 16 bit and the further the value should be bigger in your program.

qq664956261 commented 5 years ago

Does your CameraTrajectory.txt use kitti groudtruth? Why do I have this warning：TF to MSG: Quaternion Not Properly Normalized？ @speed8928 @shichaoy

shichaoy commented 5 years ago

@qq664956261 It is the same format as kitti ground truth. each row is 12 number, a flattened vector of [R, t] matrix in row order. actually we get the CameraTrajectory.txt from orb slam. Do you use the provided trajectory?

qq664956261 commented 5 years ago

thanks your sharing replying and the problem has been solved.but i have another problem. why the label color is wrong the car is green and the road is purple 2019-04-18 09-48-00屏幕截图 @shichaoy

speed8928 commented 5 years ago

This is because you did not change your labels to correspond to their crf model. Have a look in data process folder.

speed8928 commented 5 years ago

@shichaoy @YiheChen1995 This might be not the question fit for the topic but do you guys know if there are any paper regarding estimate objects (static) in the semantic scene by using a similar fashion (stereo) as you did?

shichaoy commented 5 years ago

@speed8928 sorry, I didn't keep up to date in the field, but I think it is easy to find them. Maybe start with some representative work then find which paper cites them, for example "Incremental dense semantic stereo fusion ..." by Vinnet. You can also find which paper cites my work.