[HN Gopher] Histogramo - Dice Statistics on Demand
___________________________________________________________________
Histogramo - Dice Statistics on Demand
Author : DanielBank
Score : 4 points
Date : 2022-05-21 18:12 UTC (4 hours ago)
(HTM) web link (www.histogramo.com)
(TXT) w3m dump (www.histogramo.com)
| JellyBeanThief wrote:
| Yep, that sure is a nice icon.
| DanielBank wrote:
| I liked it too. It was designed by
| https://www.fiverr.com/design_desk/
| gus_massa wrote:
| The landing page has too little information, so I'd copy the
| description in the Google store for you:
|
| > _Detect dice rolls with your camera and track the statistics of
| your dice games. Histogramo runs offline: no internet connection
| is required and camera data stays on your device. Please consider
| buying Histogramo Pro for the ability to save historical game
| data._
|
| Here sometimes it's easier to get traction with a technical post.
| Do you have an interesting technical anecdote to write about?
|
| Which model of dice are harder to detect? White or transparent?
| White dice on a white table? Small single dot for 1 or big single
| dot or custom drawing? L or R dices? Do you detect only D6, or
| also D4, D8, D12 and D20?
| DanielBank wrote:
| Thank you for the advice (I'm new to this). The model performs
| better with a dark background, for example white dice on a wood
| table. I trained it with different colors of dice in the hopes
| that that would help it generalize. The model also performs
| better with top-down images of dice. I did train with a lot of
| oblique angles of dice (3000 images in total). That helped
| somewhat but it is still far from ideal. The model currently
| recognizes a standard D6 dice with pips. I made a separate
| model that recognizes D6 and the boat / castle die from Catan:
| Cities and Knights. That model is only used with the associated
| game type.
|
| Classifying dice rolls is an interesting problem to me because
| the same "orientation" of the die might translate into
| different results of what is the actual value. The model has no
| concept of what "up" is. The oblique angles issue also adds
| complexity.
|
| In my current approach, I opted to train the model to recognize
| an entire die. Another thought I had was to train it to
| recognize the individual faces as well as the tabletop. Then if
| I could find the face that was furthest away from the tabletop,
| I could predict that as the value.
|
| One last problem which I haven't solved yet is the concept of
| re-rolls. The detection is somewhat simplistic in that it looks
| for N consecutive rolls with the same result. It requires an
| empty frame where nothing is detected to "prime" the detector
| for the next roll. This was done because I didn't want the
| detector to redundantly record the same image of dice as
| multiple rolls.
|
| I drew a lot of inspiration for this project from the following
| tutorial: https://github.com/nell-byler/dice_detection. Nell
| also has a Dice Dataset on Kaggle:
| https://www.kaggle.com/nellbyler/d6-dice
| gus_massa wrote:
| As the other comment says, you should copy the description in
| the landing page, and also add a few screenshots. It
| essentially duplicates the information in the page is Google
| Play, but not everybody clicks in the bottom to go to Google
| Play. Many (most?) persons just click the back button and
| forget your project.
| DanielBank wrote:
| Thank you again! I updated the website to at least have the
| hook from Google Play and the screenshots.
| jonnycomputer wrote:
| Not everyone reading HN is on a phone, or is going to be willing
| to rush to install an app. Sell it, man.
___________________________________________________________________
(page generated 2022-05-21 23:02 UTC)