2025/03/21 Weekly Meeting Notes

[himanshunaidu]

Apart from adding updates to the Scenario 13 of ORDP, here are the following updates for iOSPointMapper from Himanshu's side:

1. Computer Vision Model Update
   As you pointed out, we currently do not have the sufficient quantitative evaluation metrics for the models that we are using.
   Hence, I implemented some of the relevant evaluation metrics in the ML pipeline that I am developing. Some of these metrics are commonly used metrics such as mIoU, while I also went through Ricky's paper and implemented a version of ROM and RUM.
   I will soon provide some evaluation results for the models that I am working on for the app.

2. Panoptic Segmentation
   Also, as pointed out, using instance segmentation methods like the watershed algorithm may be a starting point, but empirically, the methods used in OASIS framework seem to work better by utilizing multiple segmentation masks.
   While it seems possible to implement this on iOS using OpenCV C++, I am facing some unusual issues. Some of the C++ code I am writing is not being recognized by the compiler, and I haven't been able to figure out why. I have workarounds in mind, but they may add a lot of code debt, hence I am spending more time debugging this.

3. Segmentation (Sidewalk and other line segments) Post-Processing
   As Nik mentioned, he will not be available to work on the project, hence I am picking up the work for the line segment post-processing.
   This is also taking a while because I plan to implement this using OpenCV C++, so this also depends on the problem mentioned in "Panoptic Segmentation".

4. Image Annotation
   Figuring out how to efficiently work with Scalabel instead of LabelMe. Will start actual image annotation once the above issues are taken care of.