OMSCS CS6476 (Computer Vision) Review and Tips
Eugene Yan
A grad student wrote up his deep dive into Georgia Tech's online Computer Vision course, CS6476. The surprising bit: he says online learning beat his on-campus undergrad experience, hands down.
Based on reporting by Eugene Yan — read the original for the full story.
Summary, retelling and take written by AI under human oversight; images are AI-generated illustrations. How we work · Report an error
Eugene Yan just finished CS6476, the computer vision course in Georgia Tech's OMSCS program, and his write-up reads less like a syllabus review and more like a field report from someone who got humbled by Numpy. He'd already built image classification and reverse image search APIs using deep learning libraries before enrolling. What he wanted was the plumbing underneath — the actual math of how images behave as data, not just another Keras tutorial.
The course delivers that in a fairly brutal way. Three hours of lecture a week, an assignment every other week, and each assignment split into building the algorithm from scratch in Python and then running experiments to tune it for a report. Students implemented Hough transforms to count coins by detecting circles, stereo matching to pull depth from multiple images, Harris and SIFT with RANSAC for stitching panoramas, optical flow for tracking motion, and particle filters for following subjects across video frames. All of it in raw Numpy, with an autograder enforcing strict timeouts — loop-heavy code simply failed, which forced everyone into vectorized thinking whether they wanted it or not.
Yan clocked 15 to 20 hours a week on some problem sets, which is a lot for a part-time online program. But the thing he keeps returning to is how wrong his assumptions about online education turned out to be. Recorded lectures, an active Piazza forum, recorded office hours, and a Slack channel for griping through tough assignments added up to more peer learning than he says he got as an undergrad sitting in a real classroom. That's a small but pointed data point in the ongoing argument about whether remote learning is a downgrade or just a different shape of the same thing.
The bigger payoff, he argues, wasn't any single algorithm but the translation skill — turning research papers and equations into code that actually runs, and noticing where lab-clean results fall apart in messy real conditions. He name-checks a deep learning conference demo of fall-detection video tracking as a moment where that intuition let him spot the technical seams before the presenter explained them. And he's already planning to point the skills at Lazada's image pipeline: deduplication, cleanup, classification, the unglamorous work that eats a lot of engineering time at any company sitting on millions of user-uploaded photos.
My take — AI-written commentary, not fact-checked reporting
I like seeing someone go back to first principles instead of stacking another pretrained model on top of a problem they don't understand — that's the kind of grind most people skip now that vision APIs are one API call away. The online-versus-campus point matters too: async video plus a good forum plus TAs who actually show up beats a lecture hall more often than universities like to admit, and OMSCS has been proving that for years while plenty of prestige programs still pretend otherwise.
Read more about this at: Eugene Yan