Missing data for your classifier? No problem - just let AI generate some for you! 🧠
Nearly five years ago, Eugene and I submitted our master's thesis where we used GAN-generated images of people for training data. This weekend, I revisited this idea and discovered that several scientific papers have since adopted similar approaches.
One such example is a Pet Classifier that utilizes GAN images for data augmentation: GAN-Augmented-Pet-Classifier by mahd-darvish.
Why use GANs for data augmentation, and in which cases does it make sense? 🤔
As the authors point out, augmenting datasets with GAN-generated images increases diversity by introducing more variations in pose, scale, lighting, and object positions. This enhancement helps improve the model's ability to detect and classify fine-grained differences more effectively.
So, if you're facing underfitting or overfitting issues due to a lack of variation in your dataset, generated images might be the solution. Another example, some researchers collects images from zoos to build animal classifiers due to no open-source dataset is available, could potentially benefit from GAN-generated images before testing on real animals.
But let's address the elephant in the room (pun intended). 🐘 While GANs open up many new doors in data augmentation, real data is always better. Real images contain more information about animal behavior and specific identity markings, which are crucial for re-identification and other nuanced analyses. Moreover, GANs are power-hungry algorithms, and in the context of a green agenda, the electricity consumption should be weighed against the benefits.
For this experiment, I used DALL·E with the following prompt:
"Create a realistic wildlife image, taken from a trail camera. The aspect ratio should be 16:9. Ensure the environment and animal look as realistic as possible. Include a random number (1-5) of the animal at different distances from the 'camera.' In the footer, where a trail camera typically displays information, add the text 'Animal Detect' in one corner and 'by Hugo Markoff' in the other. Start with the animal: Aardwolf."
(You can see that some images are far from what I asked for)
I then processed the images through EcoAssist by Addax Data Science, where the MegaDetector was used for detection and the Namibian Desert classifier model for classification. Some images were correctly classified with high confidence, while others were less successful. Whether this is due to the classifier or the GAN-generated images, I'm not certain.
What's your take? Can you guess which animals each image is supposed to represent? 🦁🐆
Looking forward to hearing your thoughts!
