The task was to create a Face-vs-Place classification algorithm. Arguably best approach for this task was transfer-learning. Google provides a nice script for this:
https://github.com/tensorflow/tensorflow/tree/master/tensorflow/examples/image_retraining
State of the art Inception V3 was used as a foundation Then its last layer was removed and the network was re-trained on faces/places dataset. Dataset was combined from different sources.
- http://pics.psych.stir.ac.uk/2D_face_sets.htm
- http://www.vision.caltech.edu/archive.html
- http://vis-www.cs.umass.edu/lfw/
- http://vintage.winklerbros.net/facescrub.html
Only 4000 images from each dataset were kept. After some runs of the algorithm, some misclassified images from "places" data set stood out. After inspection, it became clear that many images in that dataset included faces. Some of them were manually removed before algorithm was re-trained.
Here're final stats:
INFO:tensorflow:2017-07-22 14:20:59.116129: Step 9999: Train accuracy = 100.0%
INFO:tensorflow:2017-07-22 14:20:59.116309: Step 9999: Cross entropy = 0.013106
INFO:tensorflow:2017-07-22 14:20:59.841267: Step 9999: Validation accuracy = 97.7% (N=1000)
INFO:tensorflow:Final test accuracy = 98.3% (N=300)
Also, following image was classifies as "place", while it came from a "face" dataset:
A small trick to make sure images with relatively smaller faces are correctly identified was used. The network is run twice, first on the original image, and then on the cropped centre of the image.
The easiest way to run Face-vs-Place is by using Docker image:
docker run -it -p 8080:8080 avloss/faceorplace
Then navigate to:
It's also possible to pass an image location to the script, for instance using curl:
curl http://localhost:8080/file?file_name=/__PATH__/__TO__/__FILE__