Printable adversarial T-shirt pattern with webcam detector demo

GOAL: I want to reproduce published adversarial clothing research (Thys et al. MOTION: 2019 "Fooling automated surveillance cameras", Wu et al. "Making an Invisibility Cloak") on my own PC and webcam, to test how robust open-source person detectors are. Goal: generate a printable pattern that lowers the "person" confidence of an open-source detector when I hold it in front of my torso or wear it on a shirt. step 1 → set up Python + PyTorch + Ultralytics. Target: a small current YOLO, person class only. Use my GPU if available and verify it works first. Measure the clean detection rate on test images with no patch. step 2 → data: a COCO person subset (a few thousand images) or INRIA Person, with a held-out test split I never train on. Patch: a learnable 300x300 RGB tensor from random noise. Loss = max "person" confidence on the patched boxes (use the raw model output before NMS so gradients flow) + total variation. Train with Adam, save the patch every epoch, log mean person confidence: no patch vs random-noise patch vs my patch. If mine isn't clearly better than random, say so. step 3 → robustness: paste the patch on each person's torso with random scale, rotation ±20°, perspective, brightness/contrast, blur, noise and wrinkle-like distortion. step 4 → printability: generate a color chart for me to print and photograph, build the palette from that photo, add a non-printability score to the loss. step 5 → evaluate on the held-out split: % of patched persons NOT detected at conf 0.25, vs the clean rate. Then test transfer on a second detector (e.g. RT-DETR) and report it honestly. step 6 → export a print-ready PDF (A3, 300 dpi) and a tiled T-shirt version. step 7 → live demo: webcam + detector, person boxes with confidence, live confidence graph, FPS, dashboard overlay. Record to mp4.