00/04 · INTRO

SAM VS YOLO

Transfer learning for Super Tux Kart

setup

┌──────────────────────────────────────────┐
  martin@omarchy
  --------------------
  OS       Omarchy 4.0.2-1 (Arch Linux)
  Kernel   Linux 7.1.9-arch1-2
  WM       Hyprland 0.56.2 (Wayland)
  CPU      Intel Core i7-14700KF (16+12) @ 5.60GHz
  GPU      NVIDIA GeForce RTX 4060 Ti
  RAM      31.15 GiB
  Disk     929.50 GiB btrfs
└──────────────────────────────────────────┘

results

modelbackboneresepochspix. acc.mIoUlatencyfps
U-Net (baseline)custom128²15~88.0%~0.42~8.5ms~118
YOLOv8n transferCSPDarknet400²592.86%0.562~1.0ms>1000
SAM 2.1 transferHiera-Tiny400²595.28%0.620~27.8ms~36

SAM 2.1 wins on fidelity (fine boundaries, small objects); YOLOv8n wins on throughput by ~30x.

live demos

SAM 2.1 — live segmentation

open in drive ↗

YOLOv8n — live inference

open in drive ↗

get the code / weights