Blog

How old are you?

April 5, 2017

Let’s face it, most people have never heard of Sighthound. Yet every day we get potential customers writing to us asking about some aspect of computer vision that they need for a product or service. Sometimes it’s not a good fit, but sometimes it is. They want facial recognition, or a deep network running on an iPhone, or a car recognizer, or something as simple as an age and gender recognizer. And the answer in those cases is, yes, ours is very good indeed.

Take our latest age models. We don’t put out a press release every time we release new computer vision models. “Best Age Model Now Even Better!” doesn’t tend to light up the wires. But here’s a table of mean absolute error (MAE) for age detection. Sighthound tops the list, as you’d expect from something I’ve gone to the trouble of highlighting in a blog post.

Sighthound Dataset consists of 3929 images of 81 age categories (10 to 90)
Methods MAE (Mean Absolute Error)
Sighthound 5.76
Rothe et al. [1] 7.34
Microsoft [2] 7.62
Face++ [3] 11.04

But now look at this table. This has the best academic papers for the last three years, including one likely to appear at the main computer vision conference this year – CVPR 2017 in Hawaii in July.

Group Dataset [1]. 7 groups of age. 3500 images for training & 1050 for testing
Methods Accuracy (top1) Accuracy (1-off)
Sighthound 70.5% 96.2%
Hou et al. [4] 65.0% 96.1%
Rothe et al. [1] 62.3% 94.3%
Dong et al. [5] 56.0% 92.0%
Gallegher et al. [6] 42.9% 78.1%

See bottom for references.

So this is as leading-edge as you can get in the academic computer vision world. And yet Sighthound beats them. We sometimes get universities asking to use our software as part of their research work, perhaps to provide a more accurate component to some larger project, which we’re normally happy to agree to. But it still sometimes surprises me that our team does so well against the Googles and Microsofts of the world. I suppose it shouldn’t.

The age API is live in the Sighthound Cloud. We’re working on ways to make it more accessible to non-developers. In the meantime upload some images and see how we do.

Note: The results for Face++ and Microsoft are obtained using their cloud API on October 2016.

References

[1]. Hou, Le, Chen-Ping Yu, and Dimitris Samaras. "Squared Earth Mover's Distance-based Loss for Training Deep Neural Networks." arXiv preprint arXiv:1611.05916 (2016)

[2]. DEX: Deep EXpectation of apparent age from a single image, Rasmus Rothe, Radu Timofte, Luc Van Gool, International Conference on Computer Vision (ICCV), 2015 , *Winner of LAP challenge on apparent age estimation, ∗ NVIDIA ChaLearn LAP 2015 Best Paper Award

[3]. Dong, Yuan, Yinan Liu, and Shiguo Lian. "Automatic age estimation based on deep learning algorithm." Neurocomputing 187 (2016): 4-10.

[4]. Gallagher, Andrew C., and Tsuhan Chen. "Understanding images of groups of people." Computer Vision and Pattern Recognition, 2009. CVPR 2009. IEEE Conference on. IEEE, 2009.