Explore ImageNet, ImageNet sample images, and COCO-dataset. Read and reflect on Excavating AI: The Politics of Images in Machine Learning Training Sets by Kate Crawford and Trevor Paglen and Humans of AI by Philipp Schmitt.
Although I had previously read *Excavating AI: The Politics of Images in Machine Learning Training Sets, I was reminded **of the nuanced and complicated processes that make up human perception, and the inherently political undertaking that labeling large image-based datasets becomes. I had hoped ***Humans of AI, may propose alternative solutions to the somewhat intractable nature of image labeling, but instead it reified a complex problem with few proposed solutions to inspire or incite progressive action. Perhaps because I now feel keenly aware of the ethical and political implications these training processes pose, I sometimes feel defeated by the lack of actionable recommendations that follow some of these very justified, and well researched concerns.
Using the code examples above, try running image classification or object detection on a variety of images or video What do the models recognize properly? What do they not recognize? What other aspects of the image affect the classification, including but not limited to position, scale, lighting, etc.
I’m working on a research project related to my family’s Belgian ancestry, and thought it might be interesting to test some of the projects photos using the Drag and Drop Image Classification sketch. The results were intriguing - I initially researched unfamiliar labels like "buckler," "Windsor tie," and "pickelhaube" through Kagi and Google to cross reference the resemblance of a label to my input image. Then I searched the ImageNet sample images, for the same keywords and added the results in a toggle below each image⬇️. While not all terms had corresponding examples, some results clarified the labeling logic, especially for images like the ship, labeled “wreck” and my grandfather's mugshot-style photo labeled, "Windsor tie."
