As described in #38 , rastervision is a fairly heavy dependency since it in turn pulls in a variety of complex dependencies including pytorch. In addition to this, I found it fairly unwieldy to work with and in some cases, completely broken.
Raster Vision was required in two parts of the code: for doing tiled orthomosaic workflows and for computing evaluation metrics. Both of these have been replaced with implementations directly using geospatial libraries such as rasterio, geopandas, and shapely, which were already core dependencies of this project. The one functionality that was removed is computing an exhaustive set of metrics from a confusion matrix, but this could be re-implemented at a later date.
As described in #38 ,
rastervision
is a fairly heavy dependency since it in turn pulls in a variety of complex dependencies includingpytorch
. In addition to this, I found it fairly unwieldy to work with and in some cases, completely broken.Raster Vision was required in two parts of the code: for doing tiled orthomosaic workflows and for computing evaluation metrics. Both of these have been replaced with implementations directly using geospatial libraries such as
rasterio
,geopandas
, andshapely
, which were already core dependencies of this project. The one functionality that was removed is computing an exhaustive set of metrics from a confusion matrix, but this could be re-implemented at a later date.