today is a good day.
stitching errors halved compared to yesterday, no EBs outside of borders. and restitch was again, almost perfect.
one big discovery. what i tried to tune before turned out to be the last thing i should’ve touched. both instance threshold and text prompt were untouched today because pixel level thresholding was THE variable impeding my operational efficiency.
ofc it still has to pair with hull alpha adjustmets due to the nature of this thresholding technique being “pixel-level.” but it’s fantastic nonetheless, rendering both instance threshold and text prompt the baselines that require no adjustments.
the ui is still buggy and according to cc no one can manipulate the smoothness of matlibplot that is a layer (Python) on top of macOS. it is what it is.
the long term challenge that should remain being disregarded as a concern of today is the fine tuning problem. the erratic behaviors of the native sam3-mlx, with timing and extra steps such as CCL and a universal hull alpha could, under the assumption that the PoC result is fruitful with my current intervention, adequately improve the performance of a prospective fine-tuned, domain specific model for segmentation!