sahandv
commented
1 year ago
- These scripts are clunky and probably riddled with issues, however they might be a good starting point to understand the problem.
This is absolutely a great starting point!
- [ ] Review gis2coco
I guess we can use a very tiny sample raster to better grasp the files and the way they work.
- [ ] Identify avenues for possible improvements and post in response to this issue here
I didn't have a sample to run, but here are my initial discussion points so far:
- The first thing is that we need is a significant documentation overhaul of this repo.
- Is geojson the chosen format for output shape/meta/data files? Or we should also consider things like shape files? I personally think geojson is easier to work with and parse, but on the other hand, we may use Shapely library a lot and I am not sure how well it is supported there.
- pixel 2 coord and vice versa is a thing to be careful about. I can see that GDAL is used in the existing code. (Or I am confusing things in the code maybe :) ) Was wondering if we can directly use rasterio (with GDAL backend) instead of GDAL directly. I am leaving this link here for future me. This should work both ways.
Overall a group discussion (with @mauch ) is needed to further dissect this code and build upon it, possibly in the coming week.
Last year I created some scripts to take input raster imagery layers and polygon shapefiles and convert them into COCO datasets with many smaller image tiles and associated annotations in COCO JSON.
These scripts are clunky and probably riddled with issues, however they might be a good starting point to understand the problem.