there was a blizzard that started yesterday close to today midnight, and sustained until 7 am. i woke up quite late today and decided today's primary workouts would be my roundtrip walk b/w amherst to lsl.
by 2pm, the snow has reduced its intensity and it had became relatively safe for outdoor walking (pvta was shut down for today) as long as i put on sam's ski goggle... which i did, and headed my way.
on the walk, i started playing death stranding tracks and tracks w/ similar style. in the middle of the snow and some ground hasn't been shoveled completely, with a goggle and my backpack and my watch being the main GUI to navigate directions, i felt like sam porter bridges treading around mountain knot city.
one thing i have done is after experiencing the right strap of my backpack continously attempt to slip off from my right shoulder, i realized the problem is the length imbalance of the two sides of the shoulder straps. i adjusted them and it became much tighter to my body, which helped reminding me constantly to walk with a more straightened back.
5 observations today:
- the orientation of the plate as i put into the machine was the opposite and i quickly fixed it before imaging started. this requires extra attention from now on. extremely critical and failure on this matter would be extremely embarassing and tragically bad.
- 2 of the col 12 EBs' positions are, although incrementally, but partially outside of the imaging machine's peripheral, making the mask 100% inaccurate. in short, they touched the border, and it wasn't worth to risk to do any physical shaking of the plate.
- the native stitches are becoming increasingly erratic and faulty. the faulty stitches jumped from an avg of 2-3 from just 3 days earlier to now 22. the sitch repair worked 98% of the time, but this trend indicates that the reliance on the stitch repair pipeline would increase progressively from now to at least day 21. day 22 onwards could be a different story.
- the segmentation pipeline with my current toolbox (vanilla sam3-mlx w/ instance lvl thresholding + text prompting + hull alpha adjustments) are becoming increasingly ineffective and crude against images in general, especially the stitches that are (either should or shouldn't be) showing halo effects on their borders. to expand my toolbox would be involving the increase of things i should be able to do with sam3-mlx, such as pixel level thresholding, a tunable GUI with all the current existing adjustments mid GUI run (to reduce time and make everything much faster; eg sliders for all the thresholding and hull techniques, repromptable text fields, box prompting capabilities, undo/redo buttons), and a safety pin for keyboard shortcuts while i add the re-prompt (text) feature directly inside the GUI.
all masking actions can be undone and redone using the review approval pipeline. and any free time would give me the opportunity to rerun the entire thing, now with much bigger toolbox with currently the most capable segmentation model i can use for this specific purpose (aka sam3-mlx). i have medium-to-high expectations for my segmentation pipeline adjustments.
the inherent noise that cannot be avoided is the second issue (EB partially outside of frame). everything else could be improved and approximate closer to truth. ik this is a silly thing i said but at this moment, there is potential for approximation to at least get better, either marginally or significantly.