Scale wrong problem

[oym1994]

Hello, thanks for your contribution for this edge-based slam. It's really a good innovation to use such a kind of feature in slam. I have tested it on openloris market dataset. From the result, it seems that RESLAM outputs a wrong scale. But as I have known, depth data could help recover the scale. Maybe I have set a wrong config(Maybe there's is a parameter for this), or RESLAM really has the scale problem? Thank you, really hope to get your response!

Best regards, Slamer

[fabianschenk]

Hi, thanks for trying out RESLAM. I'm always happy when people run my stuff. It looks like, you're off by a factor of ~5, which comes indeed from a wrong value in the config file. There's a setting: InputDepthScaleFactor: XXXX, e.g. dataset_tum1.yaml, which converts the values in the depth map to meters, i.e. you divide the value in the depth map by this factor. For the TUM dataset, the conversion factor is 5000.0 but for many other datasets and most sensors you probably have to use InputDepthScaleFactor: 1000.0. Just try to set it to InputDepthScaleFactor: 1000.0 and rerun. I hope this helps :crossed_fingers:

[oym1994]

Hi, thanks for ur response and sorry for my late response. I adjusted the parameter and now it seems the scale is ok. But it consumes to much cpu and memory resource of my laptop which makes it not able to finish running on the whole dataset.

The hardware is as below:

Intel i7-8750 2.2GHz 16G Memory

The executable is generated in release mode. The image resolution is 848*480

But my friend successfully run RESLAM in a stm32 chip in real time.

Besides, may u tell us the reason why u decided not to research deep on this project? I really envy the idea of edge based slam.

Best regards Slamer

[fabianschenk]

Hi Slamer,

Strange that it consumes so much memory because I discard all the non-keyframes and store only the edges and distance transform of the keyframes. However, if you have many KFs, this might still consume quite a lot of memory. You could try to run RESLAM without the Pangolin visualization, since the GUI also needs quite some memory, which can be problematic when you don't have a dedicated GPU.

Regarding why u decided not to research deep on this project: After the RESLAM paper, I finished my PhD and left university so the priorities changed quickly and I don't have time to research/maintain this project in my free-time. There are for sure still many research opportunities for edge-based SLAM, ranging from classical computer vision to deep learning. The code is also far from perfect and I released it to facilitate research for other people.

[oym1994]

Got it! Thanks for ur kind response!