
John Mark Wendler
2K posts



The joys of Computer Vision. Or: "Competitive MS Paint"

what’s the catch with SAM models from Meta that no one seems to be using them to build the obvious awesome products one could build on top of them? I don’t get it

we did that! "seam" is not picked up well as you might imagine haha so we fine tuned a few thousand images for that, but "ball" mAP50-95 is saturated out of the box on sam3 YOLO11-seg "ball" is not very good due to how it draws contours/upsamples and never will be, but you can get to 0.8 mAP50-95 with fine tuning we use sam3 for content/visuals and yolo for quantitative work - basically yolo for latency sensitive work and then backfilled w/ sam3 yolo also obv much more deployable at the edge. though after ablation and ripping out most of the text encoder in sam3 you can get it under 3 GB VRAM pressure with quantization

going through chatgpt finance is a humbling experience





















Most "AI Trends" are just people guessing. We decided to look at the data instead. We analyzed 200,000 real world vision AI projects to see what’s actually happening on the ground. This is the first large scale analysis of its kind, offering a look at the reality of production vision deployments. No fluff. Just benchmarks that you can't find anywhere else. Read the Vision AI Trends 2026 report here: trends.roboflow.com









