Using DeepLCD in LSD-SLAM

rpng / calc

Convolutional Autoencoder for Loop Closure

BSD 3-Clause "New" or "Revised" License

190 stars 45 forks source link

Using DeepLCD in LSD-SLAM #27

Closed WuSa12956 closed 4 years ago

WuSa12956 commented 4 years ago

Hi, I have now embedded Deeplcd into LSD-SLAM for closed-loop detection. But when I run ROS, I get the error in the picture.

nmerrill67 commented 4 years ago

Cool project! If you look at my example launch file online-demo.launch, you will see the net_def and weights params. It looks like yours are not pointing to the right location. Make sure they point to where you have the .prototxt and .caffemodel files.

WuSa12956 commented 4 years ago

@nmerrill67 Hello, I use DeepLCD to filter candidate frames. The frame with the highest score is selected as the candidate frame. After that, close-loop detection is performed. In this case, do I need the online-demo example to improve my code?

The error in the picture is the error that the command ‘rosrun lsd_slam_core live_slam image:=/image_raw camera_info:=/camera_info’ appeared after I added the data set bag file. Does DeepLCD need to be compiled separately before use?

This is the code I used to refer to the.prototxt and.caffemodel file。

Many thanks

nmerrill67 commented 4 years ago

No you don't need to compile it separately. You just need to make sure you set the arguments for DeepLCD to the correct file locations wherever you instantiate the DeepLCD object. You can just look at my online demo for how to do this. You will just need to make it a ros::NodeHandle param. Making a launch file will make this easier -- especially to set other params like the GPU and whatnot. Are you familiar with how to do this with ROS? LSD SLAM has it a bit convoluted compared to normal usage, but the node handle is in IOWrapper/ROS/ROSImageStreamThread.h, where you can potentially set the param like I do here