Files
@ 29f28718a69b
Branch filter:
Location: OneEye/readme.md - annotation
29f28718a69b
2.4 KiB
text/markdown
transitional data processing
1d4ca86c9193 1d4ca86c9193 1d4ca86c9193 f1f8a2421f92 1d4ca86c9193 f1f8a2421f92 41c2b255ea2a f1f8a2421f92 f1f8a2421f92 1d4ca86c9193 f1f8a2421f92 f1f8a2421f92 f1f8a2421f92 d4478db55eec f1f8a2421f92 2ee338a32bcf f1f8a2421f92 f1f8a2421f92 f1f8a2421f92 f1f8a2421f92 1d4ca86c9193 d4478db55eec f1f8a2421f92 1d4ca86c9193 d4478db55eec f1f8a2421f92 1d4ca86c9193 f1f8a2421f92 f1f8a2421f92 1d4ca86c9193 1d4ca86c9193 f1f8a2421f92 f1f8a2421f92 1d4ca86c9193 1d4ca86c9193 1d4ca86c9193 41c2b255ea2a d4478db55eec ab6a05aec0f8 f1f8a2421f92 f1f8a2421f92 ab6a05aec0f8 d4478db55eec ab6a05aec0f8 f1f8a2421f92 f1f8a2421f92 | OneEye
======
A program to extract moves of a go game from a video and save them to SGF or broadcast them online.
The basic prototype is already working. However, it requires the user to manually mark the board. This is bothersome and more importantly, the same would be required again for any possible movement of the camera or the board. Which is unavoidable.
## Position detection ##
Therefore our current research focuses on automatic position detection. It is planned as a sequence of three stages.
### Sansa ###
First detector approximately locates the board in the picture. It uses SSD (single shot detector) with MobileNet and is working decently.
Examples (downloaded from Flickr under CC0 license):
### Dochi ###
Then the cropped and normed image is fed to the second detector, designed to return precise grid corners' coordinates. This is currently implemented by a convolutional neural network and obtaining the required precision is a major hurdle to be passed.
Pretty good:
Also good, but not enough:
### Genjo ###
With grid corners it is not difficult to rectify the image and cut it into tiles, one for each intersection. Then a third stage, Genjo classifier, decides whether the intersection is empty or occupied by a black or white stone. This is currently realized by a sample of pixels voting according to their HSI intensity and it suffices to produce satisfying results. Further improvements should be relatively easy.
## Stringing a sequence of board states to a sequence of moves ##
Base case: we have two correctly recognized positions differing by a single move (single added stone). Easy.
Issues:
* illegal positions -> ignorable
* positions unreachable from the previous state
* reachable from any past state. (Incorrect states inbetween.) How to pick the correct leaf of such a tree?
* reachable by more than one move. (Lost states inbetween.) Issues with branching factor.
* stone shifts
* stone stops being recognized -> fixable manually and even ignorable
* stone is recognized at an empty intersection. It can be occupied later for real. What do?
This is handled by the `statebag` package and its documentation also states the problem and used solution in precise terms.
|