WEBVTT 00:00.240 --> 00:01.440 -: What is Google Vision? 00:01.440 --> 00:04.500 So, this is one that a lot of people don't know about 00:04.500 --> 00:08.130 because they're really focused on Bard and DALL-E 00:08.130 --> 00:10.980 and all these kind of more famous AI tools. 00:10.980 --> 00:12.810 But the Google Vision is actually 00:12.810 --> 00:14.970 one of the first transformer models that I used, 00:14.970 --> 00:18.540 and it's pretty powerful for image recognition. 00:18.540 --> 00:20.820 It's available as an API in Google Cloud, 00:20.820 --> 00:22.440 and here's an example of it on the right, 00:22.440 --> 00:25.470 as something I did where I took an image 00:25.470 --> 00:28.470 of a woman wearing fashionable clothes 00:28.470 --> 00:31.170 and then just tagged the labels from that image. 00:31.170 --> 00:33.857 It's recognized correctly that image has a hair in it 00:33.857 --> 00:36.780 and it has a picture of her head, shoulder, eye. 00:36.780 --> 00:40.020 It's flash photography, label, her sleeve, dress. 00:40.020 --> 00:44.070 It's really useful for extracting out what's in an image. 00:44.070 --> 00:46.260 And it has more than just the label. 00:46.260 --> 00:50.430 So this is the same image uploaded to their online demo, 00:50.430 --> 00:52.980 and you can see that it extracts the objects. 00:52.980 --> 00:55.950 So it knows that there's a person in the image, for example. 00:55.950 --> 00:58.920 The face, it can detect whether there is a face 00:58.920 --> 01:03.420 and then also what the emotion of that face is. 01:03.420 --> 01:05.820 There's also a few of the other kind of properties, 01:05.820 --> 01:07.710 like dominant colors you can get, 01:07.710 --> 01:11.790 which is pretty useful I think for categorization of images. 01:11.790 --> 01:13.650 Crop hints as well, so it kind of gives you an idea 01:13.650 --> 01:17.460 of if you need to crop that image, what should you focus on? 01:17.460 --> 01:19.620 And then one which I've seen some people use 01:19.620 --> 01:21.300 is the safety thing here. 01:21.300 --> 01:24.390 This isn't tagged as an adult image, which is correct. 01:24.390 --> 01:26.910 There's no violence in here, but it is racy. 01:26.910 --> 01:28.280 It's very likely to be racy, right? 01:28.280 --> 01:30.570 So you can use this to automatically detect 01:30.570 --> 01:33.030 some not safe for work images. 01:33.030 --> 01:36.540 Specific places I've used it, one was product images, 01:36.540 --> 01:38.340 tagging what's in your product images 01:38.340 --> 01:40.770 and then being able to filter as you get 01:40.770 --> 01:42.617 a good sense of what types of product images 01:42.617 --> 01:45.750 are converting, for example, on your e-commerce website. 01:45.750 --> 01:48.540 Ad creative, this is some tagging 01:48.540 --> 01:50.250 I ran across all of the creatives 01:50.250 --> 01:51.330 in their Facebook ad account 01:51.330 --> 01:54.865 so you can see what tags tend to perform better than others 01:54.865 --> 01:58.800 and how many creatives have that tag. 01:58.800 --> 02:03.540 So, with this example, we had 23 creatives that used, 02:03.540 --> 02:05.670 that had the label of communication device 02:05.670 --> 02:07.903 or 29 with electronic device, 02:07.903 --> 02:09.810 so that's interesting in itself 02:09.810 --> 02:12.060 to see what sort of patterns are appearing. 02:12.060 --> 02:13.170 And then you can also use it 02:13.170 --> 02:14.636 for not safe for work detection. 02:14.636 --> 02:17.790 Another use case is just detecting if there are faces 02:17.790 --> 02:21.300 or detecting if there are specific labels in an image 02:21.300 --> 02:22.890 in order to correct for them. 02:22.890 --> 02:25.230 AI is not particularly good with faces, 02:25.230 --> 02:26.970 and you can use this to, 02:26.970 --> 02:29.700 after you've generated an image with DALL-E, for example, 02:29.700 --> 02:31.230 you could check if there's a face in it 02:31.230 --> 02:35.370 and then regenerate if you don't want faces in your image. 02:35.370 --> 02:37.560 Okay, Google Vision I think is really powerful. 02:37.560 --> 02:41.430 It's based on the same transformer architecture as DALL-E, 02:41.430 --> 02:44.010 but it's specifically trained 02:44.010 --> 02:48.090 for this type of labeling and entity extraction, 02:48.090 --> 02:50.583 but really useful for specific use cases.