I had an idea last month for a site based around identifying video game streamers while they are in game. If I can identify them in real time then I could have real time fantasy league style betting based around the stats... it was a fun experiment but my techniques ended up not being very practical.
In the game League of Legends I could try to find the "yellow health bar" that only your character has. Once I have the coordinates for the yellow health bar I can look above it to see the character name! Boom, with the character name I can look up stats and track all of their games automatically.
For example, watching Krepo's stream here's how I'd go about getting his character name.
from livestreamer import Livestreamer while True: session = Livestreamer() streams = session.streams('http://www.twitch.tv/%s' % streamer) if streams: stream = streams['source'] container = av.open(stream.url) video_stream = next(s for s in container.streams if s.type == b'video') image = None for packet in container.demux(video_stream): for frame in packet.decode(): image = frame.to_image() features = process_image(image)
def _find_our_champion(image, search_step=2, draw_on_image=False): # should find the yellow bar underneath our hero and then # we can use the top left of that to center our search area data = np.asarray(image) # remove any alpha channel in the data so we just have (r, g, b) data = data[:, :, :3] width, height = data.shape, data.shape character_name_coords = None # Cut off the bottom few hundred pixels, not needed height -= 200 for y in xrange(0, height): hits_this_row = 0 for x in xrange(0, width, search_step): r, g, b = data[y][x] if r > 200 and 160 < g < 230 and 30 < b < 70: # really yellow hits_this_row += 1 if hits_this_row > 20: # Top left and bottom right character_name_coords = (x - 120, y - 35, x + 100, y - 8) if draw_on_image: red = (255, 0, 0) draw = ImageDraw.Draw(image) draw.rectangle(character_name_coords, fill=red) break if character_name_coords: break return character_name_coords
def _ocr_name_box(name_box_image): # 0 means load in grayscale try: gray = cv2.cvtColor(np.array(name_box_image), 0) except TypeError: gray = cv2.cvtColor(name_box_image, 0) ret, gray = cv2.threshold(gray, 160, 255, cv2.THRESH_BINARY) # ocr code below
def _ocr_name_box(name_box_image): # processing code above gray_image = Image.fromarray(gray) return pytesseract.image_to_string(gray_image)
OCR results: Krepo
So, it worked in this particular case... this was one of the few that did work.
The idea was super fun to pursue but it doesn't perform very well for a few reasons:
So... it was a fun experiment but not very practical!
You can find all the source here.
I have been binge watching Pycon 2014 videos and I thought I might as well make that productive somehow, so here's a list of the best videos (in my very humble opinion)!
I haven't watched every single Pycon video so please be sure to recommend your favorites in the comments!
David Beazley slays it. As an expert witness he writes his own versions of simple things and comes up with interesting ways to reverse engineering/research a huge codebase.
Julie Pagano reminds us to maybe not be so hard on ourselves. This one is especially important to me, because I am starting a new job soon and there's a large part of me that feels like I may not fit in.
That is bull shit!
I don't know why we get these strange fears. Everyone has that awkward first day but I'm sure at my new job the people are looking out for my best interests. They want me to be happy and feel welcomed, so I just need to think positive and do my best—no reason to worry!
"It's Dangerous to Go Alone: Battling the Invisible Monsters in Tech - PyCon 2014"
This also goes into quite a bit of arduino stuff which is something I am trying to get into, pretty interesting.
"Cheap Helicopters In My Living Room"
Guillaume Ardaud goes into some great specific about memcached, a lot of stuff I didn't know--like you pronounce memcached "memcache dee"!
I think this guy is a pretty good speaker to model yourself after. He was confident and obviously practiced a lot beforehand, really enjoyed this presentation.
"Cache me if you can: memcached, caching patterns and best practices"
Ned Jackson Lovely did an awesome presentation on machine learning/sciki. It was very beginner friendly. My favorite part was probably the flow chart "if you have less than 50 samples, get more samples."
"Enough Machine Learning to Make Hacker News Readable Again"
Jessica McKellar talks about how we could get more kids involved with computer science. What I thought stood out about her presentation was concrete ideas/details, like: