FILTERS:

Showing posts with label A.I.. Show all posts
Showing posts with label A.I.. Show all posts

Tuesday, 22 May 2018

AI AND DARK PHOTOS

Badly lit jpeg photos repaired with nothing more than machine learning.. Jab


Watch how machine learning can enhance low-light images

Boing Boing · by Andrea James · May 22, 2018
At this year's Conference on Computer Vision and Pattern Recognition, researcher Chen Chen presented a cool project that vastly improves the quality of images captured in low-light conditions.
Imaging in low light is challenging due to low photon count and low SNR. Short-exposure images suffer from noise, while long exposure can induce blur and is often impractical. A variety of denoising, deblurring, and enhancement techniques have been proposed, but their effectiveness is limited in extreme conditions, such as video-rate imaging at night. To support the development of learning-based pipelines for low-light image processing, we introduce a dataset of raw short-exposure low-light images, with corresponding long-exposure reference images. Using the presented dataset, we develop a pipeline for processing low-light images, based on end-to-end training of a fully-convolutional network. The network operates directly on raw sensor data and replaces much of the traditional image processing pipeline, which tends to perform poorly on such data. We report promising results on the new dataset, analyze factors that affect performance, and highlight opportunities for future work.
Here's the full project page for more information.

Sunday, 20 May 2018

MACHINE LEARNING FOR KIDS

Check out this site. Some great starting points for AI/Machine learning and all aimed at kids n' classrooms. Works in Scratch. Jab


Friday, 11 May 2018

Thursday, 19 April 2018

LANDSCAPES



Following on from the character examples, here are some landscapes. These are mostly computer generated and with the last two an A.I. has copied the style of a famous artist. With a little work these places can be explored in AR/VR or normal apps. JAB









Thursday, 12 April 2018

AUTO COLOURING

This site uses an AI to convincingly colour in a sketch or line drawing. Machine learning continues to achieve things that were impossible just a few years ago. Jab





Monday, 2 April 2018

3D FROM iPHONE PHOTOS

I'm continuing to push readily available technology and the iPhone 8 is still allowing me to innovate. 

Using a single iPhone photo I was able to extract the extra info taken with the newer upgrades and use it to create a rough 3D image from a single show. 

Here you can see the original image taken by the Suzie, a depth map extracted from the info and a reconstructed 3D model in photoshop. If I had time I could make an app that does this automatically.


Jab







Friday, 2 March 2018

BETTER ENHANCEMENT

While abandoned I decided to look over some old vhs interlaced videos. I noticed that in some there were moments where for a second the scene did not change. Using a bit of wizardry I isolated some of these, and for each second I was able to extract 48 frames of video! It's interlaced so there are 2 frames in one (remember those lines?). As a test of some new methods I took frames of a friends face and tried it out. The result was quite shocking. It took a horrible grainy 600 pixel (or less) image and outputted a clear detailed 5000 pixel image of my friend. It took a while but it's fascinating that all those blurry, duplicated photos and videos we've thrown out in the past could have had a new life. I used to throw out duplicate photos not knowing that in the near future I could use the information to super sizing like this. Might be worth having for a class in the future. Jab


The image in the corner is the actual size of the original info. To see the full result CLICK HERE.


Tuesday, 13 February 2018

OUR SCHOOL EQUALISER

Continuing to push the new language to its limits and testing the new graphics and camera capabilities of Unity I created a dancing Bolt that will react LIVE to sound input. This sound can be a piece of music played on the computer or a live line in from the school's musical instruments. Even a microphone can be used.

The app uses a number of tweaked algorithms to allow bolt to react slightly differently to any beat or sound that hits him. Best part is, he doesn't need any preparation and will instantly adapt. I've given the computer control of the lights, camera and pyrotechnics and in the example video it's the computer that chooses to use them. My person favourite is the Flash Gordon music by queen. Those perfectly timed pyrotechnics and explosions are all the computer's doing.



Jab











Thursday, 8 February 2018

MACHINE LEARNS - SUZIE

Later this year Unity will gain some Machine Learning capabilities. That is, Artificial Intelligence, Neural Nets and all that jazz. As an experiment and to learn more about the subject I started to play around with machine learning. Using some off the shelf software I created a 'brain' and fed it multiple photos of Suzanne. In fact I fed it 20,000 frames of video and 850 high resolution photographs going back over the last 18 years. So no need to worry about 'faking' other people as I had to use 2 decades of memories!

At first all I was getting was blurry images but after a week, yes a whole week, the brain had studied Suzie's face so much it was producing much better results. After a few silly attempts, like changing faces in music videos I decided something more difficult.
The result is a 14 year old Suzie talking about fashion in Hebrew. To make it seem more authentic I added some VHS distortions and pitched the voice closer to Suzie's.
In Unity we should be able to use Machine Learning for many things like hand-writing recognition, voice commands, object recognition and other fun stuff. The more I looked into the subject, which I only dabbled into in the past, the more fascinating it is.
Here is the video and afterwards a frame grab of the 'original' video so you can see the face that was altered to become Suzie. Surprisingly, the brain also altered the girl's slightly awkward expressions to be more like Suzie too, the result is an overly confident teenager talking about the fashion industry.




Don't ask :)







Jab

Saturday, 20 January 2018

AI and PHOTOS

Seems like Google's version of the Adobe Sensei. Wonder if they'll actually release something any time soon. Jab


Google taught an AI to edit photos like a pro and the results are glorious

The Next Web · by Abhimanyu Ghoshal · July 14, 2017
Landscape photography is hard, no matter how beautiful an environment you’re shooting in. You need to be well-versed in composition, deal with weather conditions, know how to adjust your camera settings for the best possible shot, and then edit it to come up with a pleasing picture.
Google might be close to solving the last part of that puzzle: a couple of its Machine Perception researchers have trained a deep-learning system to identify objectively fine landscape panorama photos from Google Street View, and then artistically crop and edit them like a human photographer would. Here’s some of its handiwork:
The results don’t just speak for themselves: Google showed a bunch of these photos, along with others from various sources, and asked several pro photographers to grade them for quality; about 40 percent of Google’s submissions were perceived as being created by ‘semi-pro’- or ‘pro’-level photographers.
What’s especially interesting is that the AI is capable of applying contextually meaning adjustments in different parts of each photograph, making for dramatic lighting and more compelling images – as opposed to simply applying a filter to the entire picture or adding something predictable like a vignette.
I imagine this research will eventually find its way into Google Photos’ editing tools, or in the company’s other mobile editing app, Snapseed.
You can take a look at more AI-edited landscape pictures in this gallery, and find the research paper published by Google’s Hui Fang and Meng Zhang here.

Wednesday, 10 January 2018

THERE GOES THAT TOO

Don't need reeding comprihention anyway.

Jab


Deep neural network models score higher than humans in reading and comprehension test | KurzweilAI

kurzweilai.net · by http://helldesign.net
(credit: Alibaba Group)
Microsoft and Alibaba have developed deep neural network models that scored higher than humans in a Stanford University reading and comprehension test, Stanford Question Answering Dataset (SQuAD).
Microsoft achieved 82.650 on the ExactMatch (EM) metric* on Jan. 3, and Alibaba Group Holding Ltd. scored 82.440 on Jan. 5. The best human score so far is 82.304.
“SQuAD is a new reading comprehension dataset, consisting of questions posed by crowdworkers on a set of Wikipedia articles, where the answer to every question is a segment of text, or span, from the corresponding reading passage,” according to the Stanford NLP Group. “With 100,000+ question-answer pairs on 500+ articles, SQuAD is significantly larger than previous reading comprehension datasets.”
“A strong start to 2018 with the first model (SLQA+) to exceed human-level performance on @stanfordnlp SQuAD’s EM metric!,” said Pranav Rajpurkar, a Ph.D. student in the Stanford Machine Learning Group and lead author of a paper in Proceedings of the 2016 Conference on Empirical Methods in Natural Language Processing on SQuAD (available on open-access ArXiv). “Next challenge: the F1 metric*, where humans still lead by ~2.5 points!” (Alibaba’s SLQA+ scored 88.607 on the F1 metric and Microsoft’s r-net+ scored 88.493.)
However, challenging the “comprehension” description, Gary Marcus, PhD, a Professor of Psychology and Neural Science at NYU, notes in a tweet that “the SQUAD test shows that machines can highlight relevant passages in text, not that they understand those passages.”
“The Chinese e-commerce titan has joined the likes of Tencent Holdings Ltd.and Baidu Inc. in a race to develop AI that can enrich social media feeds, target ads and services or even aid in autonomous driving, Bloomberg notes. “Beijing has endorsed the technology in a national-level plan that calls for the country to become the industry leader 2030.”
*”The ExactMatch metric measures the percentage of predictions that match any one of the ground truth answers exactly. The F1 score metric measures the average overlap between the prediction and ground truth answer.” – Pranav Rajpurkar et al., ArXiv