Cognitive APIs. Vision. Part 3.

I have used Cognitive Services from Microsoft (part 1) and IBM Watson Services (part 2) to read my avatar image. There are two more APIs that I would like to put to the test - Google Cloud Vision API and Clarifai.

Google Cloud Vision

I already had a developer account. To use the Cloud Vision API I only had to enable it in the console and generate myself a browser key. When you sign up, Google asks for your credit card but they promise not to charge it without your permission. They also give you $300 in free trial credit and 60 days to use it.

The API itself is clearly designed for extensibility.

It’s a single endpoint that can do different things based on your request. An image can either be sent as a binary data or as a URL to a Google Storage Bucket. You can send multiple images at once and every image request can ask for different type of analysis. You can also ask for more than one type of analysis for a given image.

Google can easily add new features without adding new APIs or changing the endpoint’s semantics. Take a look:

1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
const key = '<use-your-own-key>';
const url = `https://vision.googleapis.com/v1/images:annotate?key=${key}`;

fetch(url, {
method: 'POST',
headers: new Headers({
'Content-Type': 'application/json'
}),
body: JSON.stringify({
'requests': [{
'image': {
'source': {
'gcsImageUri': 'gs://pveller/pavelveller.jpeg'
}
},
'features': [{
'type': 'LABEL_DETECTION',
'maxResults': 10
}, {
'type': 'FACE_DETECTION'
}],
}]
})
}).then(function(response) {
return response.json();
}).then(function({ responses }) {
const labels = responses[0].labelAnnotations;

console.log(labels.map((l) => `${l.description} - ${l.score.toPrecision(2)*100}%`))
});

Here’s what I got:

1
2
3
4
5
6
7
8
[ 
"hair - 95%",
"person - 94%",
"athlete - 88%",
"hairstyle - 84%",
"male - 79%",
"sports - 72%"
]

A man who definitely cares about his hair, right? :) I am not sure where the sports and athlete bits came from. I also wonder if I would get more tags (like a microphone, for example) if I could ask for features with lower scores. The API doesn’t seem to allow me to lower the threshold. I asked for ten results but got only six back.

The face detection sent down a very elaborate data structure with coordinates of all the little facial features. Things like left eye, right eye, eyebrows, nose tip, and a whole lot more. The only thing is … you can’t see the left side of my face on my avatar.

Google also tries to detect emotions. Of all that it can see - anger, joy, sorrow, surprise - none came back with anything but VERY_UNLIKELY. You can also test an image for explicit content. Same VERY_UNLIKELY for my avatar.

Very pleasant experience but I honestly expected a little more from Google’s Vision API.

I expected more because I know Google does all kinds of crazy things with deep learning in their labs. With images as of two years go and very recently with video. Maybe as those models mature, the Cloud Vision will support more features? Time will tell.

Clarifai

The easiest setup experience by far!

I was ready to go in just a few seconds, no kidding! And it also felt like the fastest response from all the APIs I tried. Very easy and intuitive to use as well:

1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
const key = '<use-your-own-key>';
const url = 'https://api.clarifai.com/v1/tag';

const data = new FormData();
data.append('url', image);
data.append('access_token', key);

fetch(url, {
method: 'POST',
body: data
}).then(function(response) {
return response.json();
}).then(function({ results }) {
const tags = results[0].result.tag;
const labels = [...tags.classes.keys()].map((i) => ({
'class': tags.classes[i],
'confidence': `${tags.probs[i].toPrecision(2)*100}%`
}));

console.log(labels.map((l) => `${l.class} - ${l.confidence}`));
});

Here’s what I got back:

1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
[
"music - 100%",
"singer - 99%",
"man - 99%",
"people - 98%",
"competition - 98%",
"musician - 98%",
"one - 98%",
"concert - 98%",
"pop - 97%",
"microphone - 97%",
"portrait - 97%",
"journalist - 96%",
"press conference - 96%",
"wear - 95%",
"administration - 95%",
"television - 94%",
"stage - 94%",
"performance - 93%",
"recreation - 92%",
"festival - 92%"
]

This is actually very close! Good feature detection with various plausible scenarios spelled out based on that. I would only question the absolute confidence in music and singer :) What about a… conference and a speaker?

Clarifai has another very interesting endpoint - Feedback. I haven’t used it but it seems that you can submit your own labels back to Clarifai and help them train and fine-tune the model. It won’t be your own classifier like Watson does. Feedback seems to be a crowdsourcing mechanism to train their main shared model(s). I only wonder how it will work without you having to specify the area of the image that each new label is attached to. In case of my avatar, conference and speaker would attach to the whole image. What about more involved images? Maybe I am missing something…


There’s a lot more computer vision APIs out there. Some are more generic and some a geared towards more specialized tasks like visual product search or logo recognition. Go give it a try!

It’s fascinating what kinds of things are just one HTTP request away.

Cognitive APIs. Vision. Part 2.

In part 1 of this blog series I had Microsoft’s Computer Vision analyze my avatar. Today I would like to ask Mr. Watson from IBM to do the same.

Setup

Same as last time, modern JavaScript and a modern browser.

Getting started with Watson APIs takes a few more steps but it’s still very intuitive. Once you’re all set with Bluemix account, you can provision the service you need and let it see your images.

API

IBM had two vision APIs. AlchemyVision has been recently merged with the Visual Recognition. If you use the original Alchemy endpoint, you will receive the following notice in the JSON response: THIS API FUNCTIONALITY IS DEPRECATED AND HAS BEEN MIGRATED TO WATSON VISUAL RECOGNITION. THIS API WILL BE DISABLED ON MAY 19, 2017.

The new unified API is a little weird. Similar to the computer vision from Microsoft, it can process binary images or can go after an image by its URL. Both need to be submitted as multipart/form-data though. Here’s an exampel from the API reference:

IBM Watson Visual Recognition API

It’s the first HTTP API that I’ve seen where I would be asked to supply JSON parameters as a file upload. You guys? Anyway. Thanks to the Blob object I can emulate multipart file upload directly from JavaScript.

Another puzzling bit is the version parameter. There’s a v3 in the URL but you also need to supply a release date of the version of the API you want to use. Trying without it gives me 400 Bad Request. There’s no version on the service instance that I provisioned so I just rolled with what’s in the API reference. It worked.

I also couldn’t use fetch with this endpoint. This time it’s not on Watson though. My browser would happily send Accept-Encoding with gzip in it and IBM’s nginx would gladly zip up the response. Chrome dev tools can deal with it but fetch apparently can’t. I get SyntaxError: Unexpected end of input at SyntaxError (native) when calling .json() on the response object.

Not sending Accept-Encoding would help but it’s one of the headers you can’t set. I had to resort to good old XHR.

And Here We Go

1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
// my avatar
const image = 'http://media.licdn.com/mpr/mpr/shrinknp_400_400/p/7/000/22b/22b/32f088c.jpg';

const key = '<use-your-own-key>';
const url = `http://gateway-a.watsonplatform.net/visual-recognition/api/v3/classify?api_key=${key}&version=2016-05-20`;

const parameters = JSON.stringify({
'url': image,
'classifier_ids': ['default'],
'owners': ['IBM'],
'threshold': 0.2
});

const data = new FormData();
data.append('parameters', new Blob([parameters], {'type': 'application/json'}));

const request = new XMLHttpRequest();

request.onload = function () {
const data = JSON.parse(request.response);
const tags = data.images[0].classifiers[0].classes;

const describe = (tag) => `${tag.class}, ${Math.round(tag.score*100)}%`;

console.log(tags.map(describe));
};

request.open('POST', url, true);
request.send(data);

The response from Watson?

Person, 100%

Yep. That’s it. That’s all the built-in classifier could tell me. You can train your own classifier(s) but they all appear to be basic. No feature detection that would allow to describe images. I tried to see all classes in the default classifier but the discovery endpoint returns 404 for default. I guess I will have to check back later ;)


I have more computer vision APIs to try. Stay tuned!

Your browser is out-of-date!

Update your browser to view this website correctly. Update my browser now

×