The Transformation of Computer Vision Over Two Decades

► Full episode: • The Role of Data in Ad...
► Listen on Spotify: open.spotify.com/episode/31I8...
► Listen on Apple Podcasts: podcasts.apple.com/ca/podcast...
►Jerome Pasquero: / jeromepasquero
►Twitter: / whats_ai
►My Newsletter (My AI updates and news clearly explained): louisbouchard.substack.com/
►Support me on Patreon: / whatsai
►Join Our AI Discord: / discord
#cv #ai #computervision

[@stevenmitchell7697]

I’ve been in computer vision for more than 20 years and in many ways, it hasn’t changed. In the late 90s we were developing active appearance models of faces and discussing the implications of synthetic faces in images and videos. Ensemble methods started coming out in the 00s and everyone was creating detectors similar to when YOLO appeared a decade and a half later. Mid 2000s CNNs started appearing in homemade implementations and I was solving OCR / Object Detection problems for a living. Only things that significantly changed with CNNs were GPUs which increased their size, ReLU, dropout, skip connections, and Python instead of C++. Shape Regression for landmarking appeared in the late 2000s and reinvented in the late 2010s using deep learning with plus and minus… slower and less accurate with faces, but able to solve more complex scenarios like hand and body pose tracking.
We were using machine learning in the past, dealt with collecting, curating, and labeling images. Only difference was most algorithms were not NN-based because the hardware and training wasn’t there, plus NN techniques were seen as old fashioned / passé until Hinton reinvigorated NNs in 2007.