Post #83

⚡️Ultralytics YOLO26 rasman chiqdi (2026-01-14)! Bu relizning markazida bitta g‘oya : edge va low-power qurilmalarda ham production’ga tez va og‘riqsiz chiqish. Men kutgan reliz😁 Nimalar o‘zgardi, asosiy yangiliklar - End-to-end, NMS-free inference: endi klassik NMS post-processing bosqichi yo‘q. Natija: latency kamroq, export va deploy jarayoni soddaroq. - CPU’da 43% gacha tezroq inference: ayniqsa edge senariylar uchun juda katta sakrash. - DFL (Distribution Focal Loss) olib tashlandi: arxitektura soddalashdi, exportability va low-power hardware bilan moslashuvchanlik yaxshilandi. - ProgLoss + STAL: trening barqarorroq, ayniqsa small object’larda aniqlikni oshirishga e'tibor qaratilgan. - MuSGD optimizer: SGD + Muon gibrid yondashuv, treningni barqaror va samaraliroq qilish uchun. - Task-specific yaxshilanishlar: - Segmentation: semantic loss + multi-scale protos - Pose: keypoint’lar uchun yuqori aniqlikda RLE - OBB: angle loss bilan boundary discontinuity muammosi ustida ishlangan. - YOLOE-26 (Open-Vocabulary): text prompt yoki visual prompt bilan zero-shot detect/segment qilish. Hatto prompt-free variantlar ham bor. YOLO26 qaysi tasklarni qamrab oladi? - Object Detection - Instance Segmentation - Image Classification - Pose Estimation - Oriented Object Detection (OBB) GitHub, YOLO26 @learncvuz