Add progressive web app config, and minor bugfix for symbol decoding #81
Conversation
The negative constant (-5) with blockSize=3 is an overfit for 8x8 cimbar based on my current (371 frame) trial set. blockSize=3 or 5 or 7 aren't *too* different if the subtraction constant is set to 0 (and intuitively maybe it *should* be). 7 seems slighly worse than the other two, and 5 (marginally) the best. When the sharpen kernel is enabled, 7 and 5 are more or less identical (3 is worse), but 7 seems best
| blockSize = 7; | ||
| sharpenSymbolGrid(symbols, symbols); | ||
| } | ||
| cv::adaptiveThreshold(symbols, symbols, 255, cv::ADAPTIVE_THRESH_MEAN_C, cv::THRESH_BINARY, blockSize, 0); |
There was a problem hiding this comment.
It's not clear the best way to calculate it (do you include good and bad images? borderline ones? etc etc), but the most sensible way I have of evaluating the error rate introduced by that -5 is to check how many of the sample images fit into certain error cutoffs.
Before and after, using blockSize=3 and different constant values (top=-5, bottom=0):
| <50 tile errors | < 100 | < 200 | < 400 | < 500 | <700 | < 1000 | total |
|---|---|---|---|---|---|---|---|
| 28 | 62 | 98 | 149 | 168 | 190 | 223 | 371 |
| 107 | 141 | 177 | 208 | 226 | 251 | 266 | 371 |
Higher numbers are better (we could also express it as a %, e.g. 141/371 == 38%), and includes all numbers to the left.
For very low error rates the difference between before and after is deceptive and mostly irrelevant (the ECC will correct 1 error or 60 with similar ease), but on the whole this means that the previous change was a significant regression under certain lighting conditions -- the 371 image dataset includes various lighting scenarios -- and fine under others.
(misc note: "tile errors" nor "bit errors" is exactly the metric we care about -- ideally we'd want to chart byte errors, since that's what reedsolomon corrects for. But tile errors is convenient to calculate, and operates as a proxy for byte errors. tile == 6 bits)
In trying to improve the color decode, I've gathered a larger dataset of real-world sample images -- and it turns out that the decoder change introduced here was overfit. Incremental learning...
The other change here is to add a
pwa.jsonfile to enable installing cimbar.org as a webapp. There are some drawbacks (I've seen some brutal performance in airplane mode on android that simply I don't understand), but in general it seems like a reasonable thing to have, even if it's imperfect.