Skip links

techesoterica.com

A blog dealing with sensory substitution and other esoteric concepts and technologies like speech-recognition and chaos theory

Main navigation

concepts

A review of Machine Learning is Fun! by Adam Geitgey

by ·

I bought the book with much anticipation since I am a regular reader of Mr. Geitgey’s posts. The book did not disappoint. I particularly enjoyed the introductory section on neural networks, specially, , the lucid description of forward and back propagation. I have read many references on the web and have taken the famous machine learning course by Andrew NG but none of those references explained how a neural network works as well as machine learning is fun did.
The code examples are easy-to-read and are well organized.
The VMware virtual machine is a nice touch.
I would have liked to see more discussion about adversarial neural networks and generative neural networks. In addition, more details of commonly used optimization algorithms such as gradient descent would have been welcome.
Finally, a section on how to install the several libraries mentioned would also be handy.

Shopping for footwear using synthetic vision

by ·

Sacheta wanted to buy footwear, so there we were, in Kala Niketan, Janpath on a Saturday evening looking for bellies and such. As usual, I was wearing my Vision 800 glasses running the vOICe. One advantage in footwear shopping is that the customer needs to sit to try it. The shop was not too crowded and we did not have a problem in finding a seat. However, before sitting, we looked around the shop, asking the staff for what she wanted. There were shelves packed with footwear. I was unable to understand the shapes until I touched a slipper. It had been stacked on its end such that the strap that comes on the top of your foot was facing the customer. This is different from how shoes are stacked because shoes are kept with their backs to the customer or at least that was the way they were kept the last time I checked.
The shapes were uniform from a far and the shelf boundaries formed hard edges in the scene. The shop had a large range of footwear in several colors. Once I was seated, I began to play with the color filter. Blue, green and red did not yield much feedback. I hit pay dirt when I chose orange. This was strange. I could hear shoppers asking for a colors like “rose gold”, whatever that is. Where was this color? I don’t think anyone wears orange shoes but then what do I know? The analyze option came in really handy. I sat back letting the scan tell me what filters were working. I then enabled the live OCR which introduced a new wrinkle. As I panned, I heard “fresh stock.” Hmm, what could they mean? Were they referring to the latest fashion? Was someone baking shoes? I could not ask Sacheta because she was engaged with a persistent sales person.
Once she had completed her shopping I was able to clear up some of the mysteries. The orange filter worked probably because the shop used a lot of yellow light. The “fresh stock” meant that there was no sale on those items. Don’t ask me why they could not say this up front.
I did have a chance to look at shoes and to distinguish details, my palm remains the best tool. However, with some more practice, I think I will be able to reliably distinguish between shelves full of slippers and shelves where shoes are stacked.
I did not buy anything so did not get a chance to try walking but I can see myself creating a visual landmark of a shelf or something else and walking independently to test shoes.
Some may ask, what is the point of it all? The simple way to buy footwear is to ask for what you want and try until you like something. Yes, this approach works well but my scene reader helped me to get a better idea of what was in the shop and thereby make a more informed choice.

Note

Photography is not allowed in the shop so I did not take any images.

Note 2

Be careful when looking and searching for a place to sit. These shops are full of humans sitting and not paying attention to their surroundings. Stay in the middle of the isle as much as you can. Let the staff guide you to a seat.

House hunting with artificial vision

by ·

I was recently looking at places to rent for my mother in-law. I was wearing the vOICe as my wife and I looked at prospective houses. For those unfamiliar with the Indian real estate market, here are a few things to note

  • Real estate is an investment for many people. Landlords construct houses specifically for renting.
  • Many houses are little more than shells with plumbing.
  • You need to check carefully because maintenance is bad here. The landlord’s idea is to get the maximum return for the lowest cost.
  • Be careful about the furniture because it can be old and unusable.
  • Real estate agents seldom give you all the facts about a place up front.

I used passive echo location to get a sense of the size of the house.
I used my nose to tell how well the air was circulating in the house which was a key to ventilation.
I did not want to touch anything because the houses were not being cleaned regularly. The vOICe however came in handy to ask about bits of furniture. I was also able to see and ask about the paint on the walls because I was able to perceive texture differences. In addition, I was able to explore independently and supplement my other senses and see the quality of doors.
I was unable to confirm if objects were dusty visually but I expect if you see something hazy, assume it is dusty.
I was also able to get an idea of how the common staircase was maintained by scanning the walls. I did feel them while ascending and descending the staircase but the vOICe gave me a larger field of view.
One example of detecting furniture was when I spotted a stack of something. It turned out to be a wall rack cut out of the slab of the open kitchen.
You may well ask me, why not ask my sighted wife to describe things to me? I could have done that but she was making her own observations, talking to the agent and navigating me. We did however had a discussion in the car where I was able to compliment some of her observations.
Our hunt continues but I am glad I have the ability to get data about my environment at a distance.

Access to travel the Planet Abled way

by ·

Planet Abled invited me to speak at the Access To Travel conference conducted by them on 27 September 2017. I went in expecting the usual conference but was pleasantly surprised. It was conducted at Park Hotel. The conference experience began just after the metal detectors. I was barely through the device when I was met by a Planet Abled team member and escorted to the venue. I checked as I entered the hall. Something was wrong. The light level dropped, and I was in an enclosed space. I am light independent therefore I continued through and entered the auditorium.It turns out that the entrance had been reconfigured as a darkness simulator.
Cameras and Planet Abled staff were everywhere. Unlike other conferences, they did not ask me to settle down. There was active focus on moving around and interaction. The staff helped. We, the tourists had a chance to meet other key leaders in the travel industry such as Mr. Subash Goyal. The food was good and was dry. I have attended more than my share of conferences where the food has gravy and is impossible to eat without 2 working hands.
it is rare to have a discussion on recreation in India. The access to travel conference was one of the few events that addressed not only the challenges when traveling with a disability but allowed me to see what everyone else was doing to have fun.
The most enjoyable part for me in the conference was the emphasis on stories. Each speaker had his own story to tell and had time to tell it in. It was also very easy to ask speakers questions and to meet those who stayed after the conference. As always, the speakers were cross disabilities. I returned home with a greater sense of unity. Our senses that we used to engage with the world were different, but the problems were the same and had the same broad solution namely making people better humans and treating each other with dignity and respect.

Artificial vision in the enterprise

by ·

I recently acquired the Vision 800 smart glasses. This gave me a compact and convenient setup allowing me to run the the vOICe. I have had several visual experiences. When I wear the glasses, I am effectively wearing an Android tablet on my head. Yes, it would be nice to do multiple things but given the specifications of the glasses, I use them as a dedicated vision device. I also use bone conduction headphones.

  • I am able to read floor numbers as well as other signage. This means that when the fire martial asks me to exit from gate 3B, I know what he is talking about. In addition, I can navigate the stairs independently and do not need to count floors.
  • I can lean forward and see if my laptop has an error message. It is easier talking to the help desk if I can tell the problem and many times, I can solve the problem independently.
  • I am better at indoor navigation. I am able to tell when silent humans are in the way.
  • The camera on the vision 800 glasses is on the extreme left. I am not used to scanning therefore the narrow field of view and the left orientation is not matching the sense of space my body has. This is taking getting used to.
  • I am also still working out the right time to look down.
  • I can derive more information about my environment such as detecting flower pots that have been placed on top of filing cabinets.
  • Bone conduction headphones are a double edged sword. Yes, I can hear environmental sounds but on the other hand, in situations like lunch time at the office Cafeteria, they are almost useless. I cannot hear the soundscapes unless I increase the volume significant in the vOICe.
  • I have run the glasses for over 8 hours. They do not heat up much.
  • I can better handle situations where colleagues leave things on different parts of my table. For example, a colleague heated my lunch and his. He kept my lunch but forgot to tell me that he had done this. I scanned the table and was able to get hold of my plate.

Post processing images including describing them automatically

by ·

As most of you know, I publish plenty of images on this blog. I ensure that all of them are described. The biggest challenge I have in posting photographs on this blog is captioning them. I have to get images described manually before I put them up here. Once I take my photographs I isolate my pictures by location because of the geotagging my phone does. I then send them to people who have been on the trip who describe the images. I have been searching for solutions that describe images automatically. I was thrilled to learn that wordpress had a plugin that used the Microsoft Cognitive Services API to automatically describe images. The describer plugin however did not give me location information therefore I rolled my own code in python. I have created a utility that queries Google for location and the Microsoft Cognitive Services API for image descriptions and writes them to a text file. I had tried to embed the descriptions in EXIF tags but that did not work and I cannot tell why.

References

You will need an API key from the below link.

Microsoft Cognative Services API

The wordpress plugin that uses the Microsoft Cognitive Services API to automatically describe images when uploading

Automatic Alternative Text

Notes

  • You will need to keep your cognitive services API key alive by describing images at least once in every 90 days I think.
  • Do account for Google’s usage limits for the geotagging API.
  • In the code, do adjust where the image files you want described live as well where you want the log file to be stored.
  • Do ensure you add your API key before you run the code.
import glob
from PIL import Image
from PIL.ExifTags import TAGS
from PIL.ExifTags import TAGS, GPSTAGS
import piexif
import requests
import json
import geocoder
def _get_if_exist(data, key):
    if key in data:
        return data[key]

    return None



def get_exif_data(fn):
    """Returns a dictionary from the exif data of an PIL Image item. Also converts the GPS Tags"""
    image = Image.open(fn)

    exif_data = {}
    info = image._getexif()
    if info:
        for tag, value in info.items():
            decoded = TAGS.get(tag, tag)
            if decoded == "GPSInfo":
                gps_data = {}
                for t in value:
                    sub_decoded = GPSTAGS.get(t, t)
                    gps_data[sub_decoded] = value[t]

                exif_data[decoded] = gps_data
            else:
                exif_data[decoded] = value

    return exif_data

def _convert_to_degrees(value):
    """Helper function to convert the GPS coordinates stored in the EXIF to degrees in float format"""
    d0 = value[0][0]
    d1 = value[0][1]
    d = float(d0) / float(d1)

    m0 = value[1][0]
    m1 = value[1][1]
    m = float(m0) / float(m1)

    s0 = value[2][0]
    s1 = value[2][1]
    s = float(s0) / float(s1)

    return d + (m / 60.0) + (s / 3600.0)

def get_lat_lon(exif_data):
    """Returns the latitude and longitude, if available, from the provided exif_data (obtained through get_exif_data above)"""
    lat = None
    lon = None

    if "GPSInfo" in exif_data:
        gps_info = exif_data["GPSInfo"]

        gps_latitude = _get_if_exist(gps_info, "GPSLatitude")
        gps_latitude_ref = _get_if_exist(gps_info, 'GPSLatitudeRef')
        gps_longitude = _get_if_exist(gps_info, 'GPSLongitude')
        gps_longitude_ref = _get_if_exist(gps_info, 'GPSLongitudeRef')

        if gps_latitude and gps_latitude_ref and gps_longitude and gps_longitude_ref:
            lat = _convert_to_degrees(gps_latitude)
            if gps_latitude_ref != "N":
                lat = 0 - lat

            lon = _convert_to_degrees(gps_longitude)
            if gps_longitude_ref != "E":
                lon = 0 - lon

    return lat, lon
def getPlaceName(fn):
    lli=()

    lli=get_lat_lon(get_exif_data(fn))
    g = geocoder.google(lli, method='reverse')
    return g.address
def getImageDescription(fn):
    payload = {'visualFeatures': 'Description'}
    files = {'file': open(fn, 'rb')}
    headers={}
    headers={ 'Ocp-Apim-Subscription-Key':     'myKey'}
    r = requests.post('https://api.projectoxford.ai/vision/v1.0/describe',         params=payload,files=files,headers=headers)
    data = json.loads(r.text)
    dscr=data['description']
    s=dscr['captions']
    s1=s[0]
    return s1['text']
def tagFile(fn,ds):
    img = Image.open(fn)
    exif_dict = piexif.load(img.info["exif"])
    exif_dict['Description''Comment']=ds
    exif_bytes = piexif.dump(exif_dict)
    piexif.insert(exif_bytes, fn)
    img.save(fn, exif=exif_bytes)
def createLog(dl):
    with open('imageDescriberLog.txt','a+') as f:
        f.write(dl)
        f.write("\n")

path = "\*.jpg"
for fname in glob.glob(path):
    print("processing:"+fname)
    createLog("processing:"+fname)
    try:
        imageLocation=getPlaceName(fname)
    except:
        createLog("error in getting location name for file: "+fname)
        pass
    try:
        imageDescription=getImageDescription(fname)
    except:
        createLog("error in getting description of file: "+fname)
        pass
    imgString="Description: "+imageDescription+"\n"+"location: "+imageLocation
    createLog(imgString)
    try:
        tagFile(fname,imgString)
    except:
        createLog("error in writing exif tag to file: "+fname)
        pass

Getting fenrir talking on a raspberry pi

by ·

Fenrir is a promising new user mode screen reader. It is primarily written in python 3. Here are my instructions to install it on a raspberry pi.
I am assuming that you are using RASPBIAN JESSIE LITE and are at a terminal prompt.

Updating your installation

You must update your installation of raspbian otherwise components like espeak will not install correctly.

sudo apt-get update

Let the update finish. The above command just fetches the listing of packages that need to be updated. Now do the actual upgrade.


sudo apt-get upgrade -y

This is going to take a while to complete.
You now have to install several dependencies for fenrir to work.

Espeak

This is the speech synthesizer fenrir will use.


sudo apt-get install libespeak-dev

Python 3


sudo apt-get install python3-pip -y

Fenrir is written in python 3.

The python daemon package


sudo apt-get install python-daemon -y

A package that allows fenrir to run as a service.

evdev


sudo pip3 install -evdev

A python package that handles keyboard input.

The speech-dispatcher package


sudo apt-get install speech-dispatcher -y

The above package is required to get fenrir to talk to a speech synthesizer such as espeak.

The ConfigParser package


sudo pip3 install configparser

You may need this dependency to parse the fenrir configuration file.

A module for checking spelling


sudo apt-get install enchant -y
sudo pip3 install pyenchant

An optional package to handle spell checking.

Git


sudo apt-get install git -y

Git is a version control system which you will use to pull down the latest source code of fenrir.

It is now time to install fenrir. Execute the following command.


git clone https://github.com/chrys87/fenrir.git

You now need to start configuring the pi for fenrir to work.

Execute the following command.


sudo spd-conf

You will be asked a series of questions. The main thing you need to change is the kind of sound output technology to use. You need to use alsa if you are using the 3.5MM headphone jack of the pi like I am doing. When you are asked about using pulseaudio type “alsa” without the quotes. Hit enter to move to the next prompt after each question. Do not forget to adjust the speed and pitch to your liking.

You now need to test your speech synthesizer configuration. Do ensure that you have your headphones or speakers ready.


sudo spd-say testing

If you hear the word “testing” you are good to start fenrir. If not, look through your logs and seek support.

To start fenrir, execute the following command.

Warning: You will need to change terminals once you execute the below command so you may want to open a new terminal and keep it handy.

Assuming you are in your home directory type the following commands.


cd fenrir/src/fenrir
sudo python3 ./fenrir

You should hear a stack of startup messages which signifies that fenrir is running.

Usage

I am still playing with fenrir therefore cannot comment much. The key bindings appear to be similar to those of the speek up screen reader. If you want to check them out, take a look at the config file in your installation directory or at the following link.

fenrir keyboard config file

There is a tutorial mode available accessed with the fenrir key and h. The fenrir key is the insert key by default.

Acknowledgements

Thanks to the following people for helping me get fenrir talking and for raspberry pi advice.

Michael A. Ray
Storm Dragon
Other members on the IRC channel called #a11y at the Server: irc.netwirc.tk; specifically:
PrinceKyle, Jeremiah, chrys and southernprince.