Possibility of rosifying the process
rohithsaro opened this issue · 1 comments
@jbehley @tano297
Hallo, thank you for your excellent work and making the code open source. Students in the field of robotic vision such as myself are always fascinated by such work. Also i have tried to contribute by helping out with the remaining open issues.
So far everything has worked, however when fed with a custom dataset, the results are not so great. Would it be possible to rosify the process so that feeding in data could be efficient so as to debug why the algorithm does not work so well with a certain dataset.
For @XuRobotics,
Hi. I have noticed your work and i was wondering if you had used the ouster raw lidar data with this algorithm, if so, how were the results for you ?
Best regards,
Rohith
sorry for the late reply.
It's certainly possible to put the code into a ros node, but we currently have no plans to perform this.
Regarding the performance on custom datasets: Unfortunately, the performance of the range image-based approaches is strongly affected by the sensor configuration (i.e., the extrinsic calibration), which can lead to very different patterns. We had also a paper on this "domain gap", where we showed that the model trained on SemanticKITTI deteriorates when applied to nuScenes (see Langer et al.) Therefore, domain adaptation approaches or different scan representations (i.e., voxel-based approaches) are maybe a way to go to close the "domain gap".
Langer et al. Domain Transfer for Semantic Segmentation of LiDAR Data using Deep Neural Networks, IROS, 2020. http://www.ipb.uni-bonn.de/pdfs/langer2020iros.pdf