Nesne alg\u0131lama, bilgisayar g\u00f6r\u00fc\u015f\u00fcn\u00fcn en temel ve kritik g\u00f6revlerinden biridir. G\u00f6r\u00fcnt\u00fc veya video ak\u0131\u015flar\u0131 i\u00e7indeki nesnelerin konumunu (s\u0131n\u0131rlay\u0131c\u0131 kutular) ve s\u0131n\u0131flar\u0131n\u0131 belirlemeyi ama\u00e7lar. Otonom ara\u00e7lardan g\u00fcvenlik sistemlerine, end\u00fcstriyel otomasyondan art\u0131r\u0131lm\u0131\u015f ger\u00e7ekli\u011fe kadar geni\u015f bir uygulama yelpazesinde merkezi bir rol oynar. Son y\u0131llarda derin \u00f6\u011frenme tekniklerinin geli\u015fimiyle birlikte nesne alg\u0131lama modelleri, hem do\u011fruluk hem de h\u0131z a\u00e7\u0131s\u0131ndan kayda de\u011fer ilerlemeler kaydetmi\u015ftir. Ancak, ger\u00e7ek zamanl\u0131 uygulamalar\u0131n artan talepleri, modellerin bu iki kritik performans metri\u011fi aras\u0131nda optimal bir denge kurmas\u0131n\u0131 gerektirmektedir.<\/p>\n
Geleneksel nesne alg\u0131lama modelleri genellikle iki ana kategoriye ayr\u0131l\u0131r: iki a\u015famal\u0131 (two-stage) ve tek a\u015famal\u0131 (one-stage) alg\u0131lay\u0131c\u0131lar. \u0130ki a\u015famal\u0131 alg\u0131lay\u0131c\u0131lar (\u00f6rne\u011fin R-CNN ailesi), y\u00fcksek do\u011fruluk sunarken, b\u00f6lge \u00f6neri a\u011flar\u0131 nedeniyle hesaplama a\u00e7\u0131s\u0131ndan yo\u011fundur ve bu da onlar\u0131 ger\u00e7ek zamanl\u0131 uygulamalar i\u00e7in yava\u015f k\u0131lar. Tek a\u015famal\u0131 alg\u0131lay\u0131c\u0131lar (\u00f6rne\u011fin YOLO ve SSD serisi), h\u0131z\u0131 art\u0131rmak i\u00e7in b\u00f6lge \u00f6neri ad\u0131m\u0131n\u0131 atlayarak do\u011frudan s\u0131n\u0131fland\u0131rma ve s\u0131n\u0131rlay\u0131c\u0131 kutu regresyonu yapar. Bu modeller ger\u00e7ek zamanl\u0131 performans sunsa da, genellikle iki a\u015famal\u0131 modellere k\u0131yasla do\u011fruluktan \u00f6d\u00fcn verirler.<\/p>\n
Transformer mimarisinin do\u011fal dil i\u015flemedeki (NLP) ba\u015far\u0131s\u0131n\u0131n ard\u0131ndan, bu mimari bilgisayar g\u00f6r\u00fc\u015f\u00fc alan\u0131na da ta\u015f\u0131nm\u0131\u015f ve nesne alg\u0131lamada devrim niteli\u011finde yenilikler getirmi\u015ftir. Facebook AI taraf\u0131ndan geli\u015ftirilen DETR (DEtection TRansformer), nesne alg\u0131lamay\u0131 u\u00e7tan uca bir dizi tahmini problemine d\u00f6n\u00fc\u015ft\u00fcrerek, geleneksel modellerdeki karma\u015f\u0131k el yap\u0131m\u0131 bile\u015fenlere (\u00f6rne\u011fin, NMS – Non-Maximum Suppression) olan ihtiyac\u0131 ortadan kald\u0131rm\u0131\u015ft\u0131r. DETR, basit ve zarif bir mimari sunsa da, yava\u015f yak\u0131nsama ve y\u00fcksek hesaplama maliyeti gibi dezavantajlara sahipti. Bu durum, DETR’\u0131n temel prensiplerini koruyarak ger\u00e7ek zamanl\u0131 performans sunabilecek daha verimli modellerin geli\u015ftirilmesi ihtiyac\u0131n\u0131 do\u011furdu.<\/p>\n
Baidu taraf\u0131ndan geli\u015ftirilen RT-DETR (Real-Time DEtection TRansformer), bu bo\u015flu\u011fu doldurmay\u0131 hedefleyen \u00f6nemli bir ad\u0131md\u0131r. RT-DETR, DETR’\u0131n u\u00e7tan uca \u00f6\u011frenme yetene\u011fini korurken, modelin verimlili\u011fini ve h\u0131z\u0131n\u0131 art\u0131rmak i\u00e7in \u00e7e\u015fitli mimari optimizasyonlar sunar. Bu makale, RT-DETR’\u0131n mimarisini, \u00e7al\u0131\u015fma prensiplerini, uygulama detaylar\u0131n\u0131 ve ger\u00e7ek zamanl\u0131 nesne alg\u0131lama projelerinde nas\u0131l kullan\u0131labilece\u011fini kapsaml\u0131 bir \u015fekilde inceleyecektir.<\/p>\n
Nesne alg\u0131lama algoritmalar\u0131, derin \u00f6\u011frenmenin y\u00fckseli\u015fiyle birlikte b\u00fcy\u00fck bir d\u00f6n\u00fc\u015f\u00fcm ge\u00e7irdi. Bu d\u00f6n\u00fc\u015f\u00fcm, modellerin hem do\u011fruluk hem de verimlilik a\u00e7\u0131s\u0131ndan s\u00fcrekli olarak geli\u015fmesini sa\u011flad\u0131.<\/p>\n
Derin \u00f6\u011frenme tabanl\u0131 nesne alg\u0131laman\u0131n ilk \u00f6nemli ad\u0131mlar\u0131, “iki a\u015famal\u0131” alg\u0131lay\u0131c\u0131larla at\u0131ld\u0131. Bu modeller, potansiyel nesne konumlar\u0131n\u0131 belirlemek i\u00e7in bir “b\u00f6lge \u00f6neri” ad\u0131m\u0131n\u0131 kullan\u0131r ve ard\u0131ndan her bir \u00f6nerilen b\u00f6lgeyi s\u0131n\u0131fland\u0131r\u0131r ve s\u0131n\u0131rlay\u0131c\u0131 kutusunu iyile\u015ftirir.<\/p>\n
* R-CNN (Region-based Convolutional Neural Network):<\/strong> Nesne alg\u0131lamada evrimsel bir ba\u015flang\u0131\u00e7 noktas\u0131yd\u0131. Se\u00e7ici arama algoritmas\u0131yla b\u00f6lge \u00f6nerileri \u00fcretir, her bir \u00f6neriyi ayr\u0131 ayr\u0131 bir CNN’den ge\u00e7irir ve ard\u0131ndan SVM ile s\u0131n\u0131fland\u0131r\u0131r. Y\u00fcksek do\u011fruluk sunsa da, her b\u00f6lge i\u00e7in ayr\u0131 CNN ge\u00e7i\u015fi nedeniyle son derece yava\u015ft\u0131. \u0130ki a\u015famal\u0131 alg\u0131lay\u0131c\u0131lar y\u00fcksek do\u011fruluk sa\u011flasa da, karma\u015f\u0131k yap\u0131lar\u0131 ve ard\u0131\u015f\u0131k i\u015flem ad\u0131mlar\u0131 nedeniyle ger\u00e7ek zamanl\u0131 uygulamalar i\u00e7in genellikle yetersiz kal\u0131r. Bu durum, “tek a\u015famal\u0131” alg\u0131lay\u0131c\u0131lar\u0131n geli\u015ftirilmesine yol a\u00e7t\u0131.<\/p>\n * YOLO (You Only Look Once):<\/strong> Nesne alg\u0131lamay\u0131 tek bir regresyon problemine d\u00f6n\u00fc\u015ft\u00fcrerek devrim yaratt\u0131. G\u00f6r\u00fcnt\u00fcy\u00fc bir \u0131zgaraya b\u00f6ler ve her \u0131zgara h\u00fccresinin do\u011frudan s\u0131n\u0131rlay\u0131c\u0131 kutular\u0131n\u0131, g\u00fcven skorlar\u0131n\u0131 ve s\u0131n\u0131f olas\u0131l\u0131klar\u0131n\u0131 tahmin etmesini sa\u011flar. Bu yakla\u015f\u0131m, son derece h\u0131zl\u0131 \u00e7\u0131kar\u0131m s\u00fcreleri sunar ancak genellikle iki a\u015famal\u0131 modellere k\u0131yasla daha d\u00fc\u015f\u00fck uzamsal do\u011fruluk ve k\u00fc\u00e7\u00fck nesneleri alg\u0131lamada zorluklar ya\u015far. Transformer mimarisinin do\u011fal dil i\u015flemede (NLP) elde etti\u011fi \u015fa\u015f\u0131rt\u0131c\u0131 ba\u015far\u0131lar, bilgisayar g\u00f6r\u00fc\u015f\u00fc toplulu\u011funun dikkatini \u00e7ekti. Bu mimari, uzun menzilli ba\u011f\u0131ml\u0131l\u0131klar\u0131 modelleme yetene\u011fi sayesinde s\u0131ral\u0131 verilerde (metin gibi) ola\u011fan\u00fcst\u00fc performans sergilemi\u015ftir. Facebook AI taraf\u0131ndan 2020’de tan\u0131t\u0131lan DETR (DEtection TRansformer), Transformer’\u0131 nesne alg\u0131lama alan\u0131na ta\u015f\u0131yan ilk model oldu.<\/p>\n DETR’\u0131n temel prensibi, nesne alg\u0131lamay\u0131 u\u00e7tan uca bir k\u00fcme tahmini problemine d\u00f6n\u00fc\u015ft\u00fcrmektir. Geleneksel modellerden farkl\u0131 olarak, DETR’\u0131n mimarisi \u015fu ana bile\u015fenlerden olu\u015fur:<\/p>\n * Omurga A\u011f\u0131 (Backbone):<\/strong> Girdi g\u00f6r\u00fcnt\u00fcs\u00fcnden bir \u00f6zellik haritas\u0131 \u00e7\u0131karmak i\u00e7in kullan\u0131lan standart bir CNN (\u00f6rne\u011fin ResNet). DETR’\u0131n en \u00f6nemli yeniliklerinden biri, “k\u00fcme e\u015fle\u015ftirme” (set prediction) yakla\u015f\u0131m\u0131d\u0131r. Model, sabit say\u0131da tahmini kutu \u00fcretir ve bu tahminleri zemin ger\u00e7eklik (ground truth) kutular\u0131yla Macar algoritmas\u0131 kullanarak en uygun \u015fekilde e\u015fle\u015ftirir. Bu e\u015fle\u015ftirme, NMS (Non-Maximum Suppression) gibi el yap\u0131m\u0131 post-processing ad\u0131mlar\u0131na olan ihtiyac\u0131 ortadan kald\u0131r\u0131r.<\/p>\n DETR’\u0131n Avantajlar\u0131:<\/strong><\/p>\n * U\u00e7tan Uca \u00d6\u011frenme:<\/strong> NMS gibi el yap\u0131m\u0131 bile\u015fenleri ortadan kald\u0131r\u0131r, bu da modeli daha basit ve daha az parametreye ba\u011f\u0131ml\u0131 hale getirir. DETR’\u0131n Dezavantajlar\u0131:<\/strong><\/p>\n * Yava\u015f Yak\u0131nsama:<\/strong> \u00d6zellikle b\u00fcy\u00fck veri setlerinde, e\u011fitim s\u0131ras\u0131nda \u00e7ok daha fazla epok gerektirir. Bu dezavantajlar, DETR’\u0131n ger\u00e7ek zamanl\u0131 uygulamalarda yayg\u0131n olarak benimsenmesini engellemi\u015ftir. Bu noktada, RT-DETR gibi modeller devreye girerek Transformer tabanl\u0131 nesne alg\u0131lamay\u0131 ger\u00e7ek zamanl\u0131 hale getirme potansiyelini sunmaktad\u0131r.<\/p>\n Baidu’nun RT-DETR modeli, DETR’\u0131n temel avantajlar\u0131n\u0131 koruyarak (u\u00e7tan uca \u00f6\u011frenme, NMS’siz tahmin) ayn\u0131 zamanda ger\u00e7ek zamanl\u0131 uygulamalar i\u00e7in gereken h\u0131z ve verimlili\u011fi sa\u011flamak amac\u0131yla geli\u015ftirilmi\u015ftir. DETR’\u0131n yava\u015f yak\u0131nsama ve y\u00fcksek hesaplama maliyeti gibi sorunlar\u0131n\u0131 hedef alarak, mimaride \u00f6nemli optimizasyonlar sunar.<\/p>\n RT-DETR’\u0131n arkas\u0131ndaki temel motivasyon, Transformer tabanl\u0131 nesne alg\u0131lay\u0131c\u0131lar\u0131n y\u00fcksek do\u011frulu\u011funu, modern tek a\u015famal\u0131 alg\u0131lay\u0131c\u0131lar\u0131n (YOLO, SSD) ger\u00e7ek zamanl\u0131 performans\u0131yla birle\u015ftirmektir. Bu hedefe ula\u015fmak i\u00e7in a\u015fa\u011f\u0131daki temel ama\u00e7lar belirlenmi\u015ftir:<\/p>\n * Verimli Omurga A\u011flar\u0131:<\/strong> Daha hafif ve daha h\u0131zl\u0131 \u00f6zellik \u00e7\u0131kar\u0131m\u0131 sa\u011flayacak omurga a\u011flar\u0131 kullanmak. RT-DETR, \u00fc\u00e7 ana bile\u015fenden olu\u015fan bir mimariye sahiptir: omurga a\u011f\u0131, hibrit kodlay\u0131c\u0131 ve Transformer kod \u00e7\u00f6z\u00fcc\u00fc.<\/p>\n * Omurga A\u011f\u0131 (Backbone):<\/strong> * Hibrit Kodlay\u0131c\u0131 (Hybrid Encoder):<\/strong> * Transformer Kod \u00c7\u00f6z\u00fcc\u00fc (Transformer Decoder):<\/strong> RT-DETR’\u0131 ger\u00e7ek zamanl\u0131 nesne alg\u0131lama i\u00e7in cazip k\u0131lan temel yenilikler \u015funlard\u0131r:<\/p>\n * Verimli Omurga A\u011flar\u0131:<\/strong> \u00d6zellikle PP-HGNet gibi optimize edilmi\u015f a\u011flar, y\u00fcksek kaliteli \u00f6zellikler \u00e7\u0131kar\u0131rken hesaplama maliyetini d\u00fc\u015f\u00fcr\u00fcr. Bu mimari yenilikler sayesinde RT-DETR, modern tek a\u015famal\u0131 alg\u0131lay\u0131c\u0131larla rekabet edebilecek d\u00fczeyde y\u00fcksek h\u0131zlarda (\u00f6rne\u011fin, 100+ FPS) \u00e7al\u0131\u015f\u0131rken, Transformer tabanl\u0131 alg\u0131lay\u0131c\u0131lar\u0131n sundu\u011fu y\u00fcksek do\u011fruluk seviyelerini (mAP) koruyabilmektedir.<\/p>\n RT-DETR modelini ger\u00e7ek zamanl\u0131 nesne alg\u0131lama projelerinde uygulamak, belirli ad\u0131mlar\u0131 ve pratik yakla\u015f\u0131mlar\u0131 gerektirir. Bu s\u00fcre\u00e7, geli\u015ftirme ortam\u0131n\u0131n kurulumundan modelin da\u011f\u0131t\u0131m\u0131na kadar uzan\u0131r.<\/p>\n Modelin uygulanmas\u0131 i\u00e7in g\u00fc\u00e7l\u00fc bir geli\u015ftirme ortam\u0131 \u015fartt\u0131r.<\/p>\n * Programlama Dili:<\/strong> Python, derin \u00f6\u011frenme projeleri i\u00e7in standart bir se\u00e7imdir. Modelin ba\u015far\u0131l\u0131 bir \u015fekilde e\u011fitilmesi i\u00e7in iyi haz\u0131rlanm\u0131\u015f bir veri seti esast\u0131r.<\/p>\n * Veri Seti Se\u00e7imi:<\/strong> COCO (Common Objects in Context) ve Pascal VOC gibi b\u00fcy\u00fck \u00f6l\u00e7ekli, genel ama\u00e7l\u0131 veri setleri, \u00f6nceden e\u011fitilmi\u015f modeller i\u00e7in veya transfer \u00f6\u011frenimi amac\u0131yla kullan\u0131labilir. \u00d6zel uygulamalar i\u00e7in kendi veri setlerinizi olu\u015fturman\u0131z gerekebilir. RT-DETR’\u0131n e\u011fitimi, di\u011fer derin \u00f6\u011frenme modellerine benzer ancak baz\u0131 \u00f6zel dikkat gerektiren ad\u0131mlar i\u00e7erir.<\/p>\n * Transfer \u00d6\u011frenimi:<\/strong> Genellikle COCO gibi b\u00fcy\u00fck veri setlerinde \u00f6nceden e\u011fitilmi\u015f RT-DETR a\u011f\u0131rl\u0131klar\u0131yla ba\u015flamak en iyi yakla\u015f\u0131md\u0131r. Bu, modelin s\u0131f\u0131rdan e\u011fitime k\u0131yasla \u00e7ok daha h\u0131zl\u0131 yak\u0131nsamas\u0131n\u0131 sa\u011flar ve daha iyi performans elde etmeye yard\u0131mc\u0131 olur. E\u011fitilmi\u015f bir RT-DETR modelini ger\u00e7ek zamanl\u0131 uygulamalarda kullanmak i\u00e7in a\u015fa\u011f\u0131daki ad\u0131mlar izlenir:<\/p>\n * Modelin Y\u00fcklenmesi:<\/strong> E\u011fitilmi\u015f model a\u011f\u0131rl\u0131klar\u0131 (checkpoint dosyalar\u0131) belle\u011fe y\u00fcklenir ve model \u00e7\u0131kar\u0131m moduna (eval mode) ge\u00e7irilir. Ger\u00e7ek zamanl\u0131 uygulamalar i\u00e7in modelin optimize edilmesi ve da\u011f\u0131t\u0131lmas\u0131 kritik \u00f6neme sahiptir.<\/p>\n * Model Nicemleme (Quantization):<\/strong> Modelin a\u011f\u0131rl\u0131klar\u0131n\u0131 ve aktivasyonlar\u0131n\u0131 daha d\u00fc\u015f\u00fck hassasiyetli formatlara (\u00f6rne\u011fin, FP32’den FP16’ya veya INT8’e) d\u00f6n\u00fc\u015ft\u00fcrmek, model boyutunu k\u00fc\u00e7\u00fclt\u00fcr ve \u00e7\u0131kar\u0131m h\u0131z\u0131n\u0131 art\u0131r\u0131r. RT-DETR’\u0131n ger\u00e7ek zamanl\u0131 nesne alg\u0131lamadaki de\u011feri, di\u011fer pop\u00fcler modellerle yap\u0131lan kar\u015f\u0131la\u015ft\u0131rmal\u0131 performans de\u011ferlendirmeleriyle ortaya konulur. Bu de\u011ferlendirmeler genellikle iki ana metrik \u00fczerine odaklan\u0131r: do\u011fruluk (mAP) ve h\u0131z (FPS).<\/p>\n RT-DETR, \u00f6zellikle h\u0131z ve do\u011fruluk aras\u0131ndaki denge noktas\u0131nda \u00f6nemli bir ilerleme kaydetmi\u015ftir. COCO veri seti \u00fczerinde yap\u0131lan standart de\u011ferlendirmelerde, RT-DETR’\u0131n farkl\u0131 boyutlardaki (\u00f6rne\u011fin, k\u00fc\u00e7\u00fck, orta, b\u00fcy\u00fck) modelleri, \u00e7e\u015fitli donan\u0131m konfig\u00fcrasyonlar\u0131nda (\u00f6rne\u011fin, NVIDIA V100, T4 GPU’lar) test edilmi\u015ftir.<\/p>\n * YOLOv7 ve YOLOv8 ile Kar\u015f\u0131la\u015ft\u0131rma:<\/strong> RT-DETR, genellikle benzer veya daha y\u00fcksek mAP de\u011ferleri sunarken, \u00f6zellikle daha y\u00fcksek \u00e7\u00f6z\u00fcn\u00fcrl\u00fcklerde ve daha b\u00fcy\u00fck modellerde YOLO serisine k\u0131yasla daha iyi FPS de\u011ferlerine ula\u015fabilmektedir. \u00d6rne\u011fin, belirli bir mAP seviyesinde, RT-DETR’\u0131n \u00e7\u0131kar\u0131m h\u0131z\u0131 YOLOv7’den %20-50 daha h\u0131zl\u0131 olabilir. Bu, Transformer mimarisinin k\u00fcresel ba\u011flam\u0131 daha iyi yakalama yetene\u011fi ve RT-DETR’\u0131n mimari optimizasyonlar\u0131n\u0131n birle\u015fimiyle a\u00e7\u0131klanabilir. Genel olarak, RT-DETR, nesne alg\u0131lama alan\u0131nda h\u0131z ve do\u011fruluk aras\u0131nda yeni bir denge noktas\u0131 belirleyerek, Transformer tabanl\u0131 mimarilerin ger\u00e7ek zamanl\u0131 uygulamalarda yayg\u0131nla\u015fmas\u0131n\u0131n \u00f6n\u00fcn\u00fc a\u00e7m\u0131\u015ft\u0131r.<\/p>\n RT-DETR, nesne alg\u0131lamada \u00f6nemli bir ad\u0131m olsa da, her teknoloji gibi kendi zorluklar\u0131n\u0131 ve gelecekteki geli\u015fim alanlar\u0131n\u0131 bar\u0131nd\u0131rmaktad\u0131r.<\/p>\n * Y\u00fcksek Hesaplama Kaynaklar\u0131 \u0130htiyac\u0131 (E\u011fitim A\u015famas\u0131nda):<\/strong> Her ne kadar RT-DETR, DETR’a g\u00f6re daha h\u0131zl\u0131 yak\u0131nsasa da, Transformer tabanl\u0131 mimarilerin do\u011fas\u0131 gere\u011fi hala \u00f6nemli miktarda GPU belle\u011fi ve hesaplama g\u00fcc\u00fc gerektirir, \u00f6zellikle b\u00fcy\u00fck veri setlerinde s\u0131f\u0131rdan e\u011fitim yap\u0131l\u0131yorsa. RT-DETR’\u0131n ba\u015far\u0131s\u0131, Transformer tabanl\u0131 nesne alg\u0131laman\u0131n gelece\u011fi i\u00e7in heyecan verici yollar a\u00e7maktad\u0131r.<\/p>\n * Daha Hafif ve Verimli Mimariler:<\/strong> Model boyutunu ve hesaplama maliyetini daha da azaltacak, ancak do\u011fruluktan \u00f6d\u00fcn vermeyecek yeni Transformer mimarilerinin geli\u015ftirilmesi devam edecektir. Mobil ve g\u00f6m\u00fcl\u00fc sistemler i\u00e7in \u00f6zel olarak tasarlanm\u0131\u015f ultra-hafif RT-DETR varyantlar\u0131 g\u00f6r\u00fclebilir. Baidu’nun RT-DETR modeli, Transformer tabanl\u0131 nesne alg\u0131laman\u0131n ger\u00e7ek zamanl\u0131 uygulamalar i\u00e7in ne kadar uygun hale getirilebilece\u011finin \u00e7arp\u0131c\u0131 bir \u00f6rne\u011fidir. DETR’\u0131n u\u00e7tan uca \u00f6\u011frenme ve NMS’siz tahmin gibi temel avantajlar\u0131n\u0131 korurken, optimize edilmi\u015f omurga a\u011flar\u0131, yenilik\u00e7i hibrit kodlay\u0131c\u0131 ve h\u0131zland\u0131r\u0131lm\u0131\u015f kod \u00e7\u00f6z\u00fcc\u00fc gibi mimari iyile\u015ftirmeler sayesinde h\u0131z ve do\u011fruluk aras\u0131nda e\u015fi benzeri g\u00f6r\u00fclmemi\u015f bir denge sunar.<\/p>\n Bu model, otonom s\u00fcr\u00fc\u015f, g\u00fcvenlik izleme, robotik ve ak\u0131ll\u0131 \u015fehir uygulamalar\u0131 gibi y\u00fcksek performansl\u0131 ve d\u00fc\u015f\u00fck gecikmeli nesne alg\u0131lama gerektiren bir\u00e7ok alanda devrim niteli\u011finde potansiyel ta\u015f\u0131maktad\u0131r. RT-DETR’\u0131n pratik uygulamas\u0131, geli\u015ftirme ortam\u0131n\u0131n do\u011fru kurulumundan veri setinin titizlikle haz\u0131rlanmas\u0131na, modelin etkin bir \u015fekilde e\u011fitilmesinden \u00e7\u0131kar\u0131m ve da\u011f\u0131t\u0131m optimizasyonlar\u0131na kadar bir dizi ad\u0131m\u0131 i\u00e7erir.<\/p>\n Kar\u015f\u0131la\u015f\u0131labilecek zorluklara ra\u011fmen, RT-DETR’\u0131n sundu\u011fu yenilikler, Transformer tabanl\u0131 mimarilerin nesne alg\u0131lama alan\u0131ndaki hakimiyetini peki\u015ftirmekte ve gelecekteki yapay zeka uygulamalar\u0131 i\u00e7in g\u00fc\u00e7l\u00fc bir temel olu\u015fturmaktad\u0131r. Modelin s\u00fcrekli geli\u015fimi, daha hafif ve verimli yap\u0131lar, daha iyi genelleme yetenekleri ve multimodal entegrasyonlar ile nesne alg\u0131lama teknolojisinin s\u0131n\u0131rlar\u0131n\u0131 zorlamaya devam edecektir. RT-DETR, ger\u00e7ek zamanl\u0131 nesne alg\u0131laman\u0131n gelece\u011fine dair umut verici bir vizyon sunmaktad\u0131r.<\/body><\/p>\n","protected":false},"excerpt":{"rendered":"Baidu’nun RT-DETR Modelini Ger\u00e7ek Zamanl\u0131 Nesne Alg\u0131lama \u0130\u00e7in Uygulama\nNesne alg\u0131lama, bilgisayar g\u00f6r\u00fc\u015f\u00fcn\u00fcn en temel ve kritik g\u00f6revlerin","protected":false},"author":1,"featured_media":0,"comment_status":"open","ping_status":"open","sticky":false,"template":"","format":"standard","meta":{"csco_page_header_type":"","csco_page_load_nextpost":"","csco_page_subscribe_form":"","csco_page_contact_form":"","footnotes":""},"categories":[1342],"tags":[],"class_list":{"0":"post-30507","1":"post","2":"type-post","3":"status-publish","4":"format-standard","6":"category-ai","7":"cs-entry","8":"cs-video-wrap"},"yoast_head":"\n
\n* Fast R-CNN:<\/strong> R-CNN’in h\u0131z sorununu \u00e7\u00f6zmek i\u00e7in t\u00fcm g\u00f6r\u00fcnt\u00fcden bir kez \u00f6zellik haritas\u0131 \u00e7\u0131kar\u0131r ve ROI (Region of Interest) pooling kullanarak b\u00f6lge \u00f6nerilerini bu harita \u00fczerinde i\u015fler. Bu, h\u0131z\u0131 \u00f6nemli \u00f6l\u00e7\u00fcde art\u0131rd\u0131 ancak b\u00f6lge \u00f6neri ad\u0131m\u0131n\u0131n hala CPU \u00fczerinde \u00e7al\u0131\u015fmas\u0131 bir darbo\u011faz olu\u015fturuyordu.
\n* Faster R-CNN:<\/strong> Fast R-CNN’deki darbo\u011faz\u0131 gideren Faster R-CNN, b\u00f6lge \u00f6neri a\u011f\u0131n\u0131 (RPN – Region Proposal Network) do\u011frudan CNN’in i\u00e7ine entegre etti. Bu, u\u00e7tan uca derin \u00f6\u011frenme tabanl\u0131 nesne alg\u0131lamay\u0131 m\u00fcmk\u00fcn k\u0131larak hem h\u0131z\u0131 hem de do\u011frulu\u011fu art\u0131rd\u0131.<\/p>\n
\n* SSD (Single Shot MultiBox Detector):<\/strong> YOLO’nun h\u0131z\u0131n\u0131 korurken do\u011frulu\u011fu art\u0131rmak i\u00e7in farkl\u0131 \u00f6l\u00e7eklerde \u00f6zellik haritalar\u0131 \u00fczerinde tahminler yapar ve \u00f6nceden tan\u0131mlanm\u0131\u015f “anchor kutular\u0131” (prior boxes) kullan\u0131r.<\/p>\nTransformer Tabanl\u0131 Nesne Alg\u0131lama: DETR<\/h4>\n
\n* Transformer Kodlay\u0131c\u0131 (Encoder):<\/strong> Omurga a\u011f\u0131ndan gelen \u00f6zellik haritas\u0131n\u0131 girdi olarak al\u0131r ve uzamsal ili\u015fkileri yakalamak i\u00e7in \u00e7ok ba\u015fl\u0131 dikkat mekanizmalar\u0131n\u0131 kullan\u0131r.
\n* Transformer Kod \u00c7\u00f6z\u00fcc\u00fc (Decoder):<\/strong> Sabit say\u0131da “nesne sorgusu” (object queries) ve kodlay\u0131c\u0131 \u00e7\u0131kt\u0131s\u0131n\u0131 girdi olarak al\u0131r. Her bir nesne sorgusu, bir nesnenin varl\u0131\u011f\u0131n\u0131 ve \u00f6zelliklerini \u00f6\u011frenmeye \u00e7al\u0131\u015f\u0131r. Kod \u00e7\u00f6z\u00fcc\u00fc, bu sorgular\u0131 kullanarak nihai s\u0131n\u0131rlay\u0131c\u0131 kutular\u0131 ve s\u0131n\u0131f etiketlerini tahmin eder.
\n* Tahmin Ba\u015fl\u0131klar\u0131 (Prediction Heads):<\/strong> Kod \u00e7\u00f6z\u00fcc\u00fc \u00e7\u0131kt\u0131lar\u0131ndan s\u0131n\u0131flar\u0131 ve kutu koordinatlar\u0131n\u0131 tahmin eden basit bir \u0130leri Beslemeli A\u011f (FFN – Feed-Forward Network).<\/p>\n
\n* Do\u011frudan K\u00fcme Tahmini:<\/strong> Do\u011frudan nesne k\u00fcmelerini tahmin eder, bu da yinelenen tahmin sorununu \u00e7\u00f6zer.
\n* K\u00fcresel Ba\u011flam Yakalama:<\/strong> Transformer’\u0131n dikkat mekanizmalar\u0131 sayesinde g\u00f6r\u00fcnt\u00fcdeki nesneler aras\u0131ndaki uzun menzilli ili\u015fkileri etkili bir \u015fekilde yakalar.<\/p>\n
\n* Y\u00fcksek Hesaplama Maliyeti:<\/strong> \u00d6zellikle y\u00fcksek \u00e7\u00f6z\u00fcn\u00fcrl\u00fckl\u00fc g\u00f6r\u00fcnt\u00fclerde ve uzun sekanslarda Transformer’\u0131n dikkat mekanizmalar\u0131 hesaplama a\u00e7\u0131s\u0131ndan pahal\u0131d\u0131r.
\n* K\u00fc\u00e7\u00fck Nesnelerde Performans:<\/strong> K\u00fc\u00e7\u00fck nesneleri alg\u0131lamada geleneksel modellere k\u0131yasla daha d\u00fc\u015f\u00fck performans sergileyebilir.<\/p>\nRT-DETR: Ger\u00e7ek Zamanl\u0131 DETR<\/h3>\n
RT-DETR’\u0131n Motivasyonu ve Hedefleri<\/h4>\n
\n* Optimize Edilmi\u015f Kodlay\u0131c\u0131:<\/strong> Transformer kodlay\u0131c\u0131n\u0131n hesaplama y\u00fck\u00fcn\u00fc azalt\u0131rken, uzamsal ba\u011flam\u0131 etkili bir \u015fekilde yakalamas\u0131n\u0131 sa\u011flamak.
\n* H\u0131zland\u0131r\u0131lm\u0131\u015f Kod \u00c7\u00f6z\u00fcc\u00fc:<\/strong> Nesne sorgular\u0131n\u0131n i\u015flenmesini ve tahminlerin \u00fcretilmesini h\u0131zland\u0131rmak.
\n* Geli\u015ftirilmi\u015f Yak\u0131nsama:<\/strong> Modelin daha az epokta y\u00fcksek performansa ula\u015fmas\u0131n\u0131 sa\u011flamak.
\n* Esneklik:<\/strong> Farkl\u0131 donan\u0131m ve performans gereksinimlerine uyum sa\u011flayabilecek \u00f6l\u00e7eklenebilir bir model sunmak.<\/p>\nRT-DETR Mimarisinin Detaylar\u0131<\/h4>\n
\n RT-DETR, verimli \u00f6zellik \u00e7\u0131kar\u0131m\u0131 i\u00e7in Baidu’nun kendi geli\u015ftirdi\u011fi PP-HGNet (PaddlePaddle Hybrid-Grid Network)<\/strong> gibi y\u00fcksek performansl\u0131 ve hafif omurga a\u011flar\u0131n\u0131 kullan\u0131r. PP-HGNet, geleneksel ResNet veya Swin Transformer gibi modellere k\u0131yasla daha az hesaplama maliyetiyle zengin ve \u00e7ok \u00f6l\u00e7ekli \u00f6zellik haritalar\u0131 \u00fcretmek \u00fczere tasarlanm\u0131\u015ft\u0131r. Bu a\u011flar, farkl\u0131 \u00e7\u00f6z\u00fcn\u00fcrl\u00fcklerde (\u00f6rne\u011fin P3, P4, P5 katmanlar\u0131) \u00f6zellik haritalar\u0131 \u00fcreterek, modelin hem k\u00fc\u00e7\u00fck hem de b\u00fcy\u00fck nesneleri alg\u0131layabilmesine olanak tan\u0131r. Verimli bir omurga a\u011f\u0131 se\u00e7imi, modelin genel h\u0131z\u0131n\u0131 ve \u00e7\u0131kar\u0131m performans\u0131n\u0131 do\u011frudan etkileyen kritik bir ad\u0131md\u0131r.<\/p>\n
\n DETR’daki standart Transformer kodlay\u0131c\u0131n\u0131n hesaplama maliyeti y\u00fcksektir, \u00f6zellikle y\u00fcksek \u00e7\u00f6z\u00fcn\u00fcrl\u00fckl\u00fc \u00f6zellik haritalar\u0131yla \u00e7al\u0131\u015f\u0131rken. RT-DETR’\u0131n hibrit kodlay\u0131c\u0131s\u0131, bu sorunu \u00e7\u00f6zmek i\u00e7in tasarlanm\u0131\u015ft\u0131r. Bu kodlay\u0131c\u0131, hem yerel (lokal) hem de k\u00fcresel (global) ba\u011flam bilgilerini etkili bir \u015fekilde birle\u015ftiren optimize edilmi\u015f bir yap\u0131 sunar.
\n * Attention-based Multi-scale Feature Fusion (Dikkat Tabanl\u0131 \u00c7ok \u00d6l\u00e7ekli \u00d6zellik Birle\u015ftirme):<\/strong> Farkl\u0131 \u00f6l\u00e7eklerdeki (P3, P4, P5) \u00f6zellik haritalar\u0131n\u0131 tek bir kodlay\u0131c\u0131da birle\u015ftirir. Bu birle\u015ftirme, basit toplama veya birle\u015ftirmeden ziyade, dikkat mekanizmalar\u0131 arac\u0131l\u0131\u011f\u0131yla yap\u0131l\u0131r. Bu sayede model, farkl\u0131 \u00f6l\u00e7eklerdeki bilgilere daha anlaml\u0131 bir \u015fekilde odaklanabilir ve nesnelerin \u00f6l\u00e7ek varyasyonlar\u0131na kar\u015f\u0131 daha sa\u011flam hale gelir.
\n * Query Selection (Sorgu Se\u00e7imi):<\/strong> Geleneksel DETR modelleri genellikle rastgele ba\u015flat\u0131lan veya \u00f6\u011frenilen nesne sorgular\u0131 kullan\u0131r. RT-DETR’\u0131n hibrit kodlay\u0131c\u0131s\u0131, daha anlaml\u0131 ve odaklanm\u0131\u015f nesne sorgular\u0131 olu\u015fturmak i\u00e7in bir “sorgu se\u00e7imi” mekanizmas\u0131 kullan\u0131r. Bu mekanizma, omurga a\u011f\u0131ndan gelen \u00f6zellik haritalar\u0131n\u0131n en belirgin veya y\u00fcksek olas\u0131l\u0131kl\u0131 b\u00f6lgelerinden sorgular\u0131 t\u00fcretebilir. \u00d6rne\u011fin, \u00f6nceden belirlenmi\u015f anchor noktalar\u0131na yak\u0131n veya y\u00fcksek g\u00fcven skoruna sahip \u00f6zellik b\u00f6lgelerinden sorgular se\u00e7ilebilir. Bu, kod \u00e7\u00f6z\u00fcc\u00fcn\u00fcn daha h\u0131zl\u0131 ve daha do\u011fru bir \u015fekilde yak\u0131nsamas\u0131n\u0131 sa\u011flar \u00e7\u00fcnk\u00fc sorgular zaten potansiyel nesne konumlar\u0131na odaklanm\u0131\u015ft\u0131r. Bu mekanizma, DETR’\u0131n yava\u015f yak\u0131nsama sorununu \u00f6nemli \u00f6l\u00e7\u00fcde hafifletir.<\/p>\n
\n RT-DETR’\u0131n kod \u00e7\u00f6z\u00fcc\u00fcs\u00fc, DETR’\u0131n temel prensiplerini korurken performans\u0131 art\u0131rmak i\u00e7in baz\u0131 optimizasyonlar i\u00e7erir.
\n * Accelerated Queries (H\u0131zland\u0131r\u0131lm\u0131\u015f Sorgular):<\/strong> Sorgu se\u00e7imi mekanizmas\u0131 sayesinde kod \u00e7\u00f6z\u00fcc\u00fc, daha anlaml\u0131 ba\u015flang\u0131\u00e7 sorgular\u0131yla \u00e7al\u0131\u015f\u0131r. Bu, kod \u00e7\u00f6z\u00fcc\u00fcn\u00fcn her bir katmanda daha az iterasyonla veya daha az say\u0131da sorguyla daha iyi sonu\u00e7lara ula\u015fmas\u0131n\u0131 sa\u011flar.
\n * Auxiliary Decoding Heads (Yard\u0131mc\u0131 Kod \u00c7\u00f6zme Ba\u015fl\u0131klar\u0131):<\/strong> Her bir kod \u00e7\u00f6z\u00fcc\u00fc katman\u0131ndan ek tahminler (s\u0131n\u0131fland\u0131rma ve kutu regresyonu) al\u0131n\u0131r. Bu yard\u0131mc\u0131 ba\u015fl\u0131klar, e\u011fitim s\u0131ras\u0131nda geri yay\u0131l\u0131m sinyalini g\u00fc\u00e7lendirerek modelin daha h\u0131zl\u0131 ve kararl\u0131 bir \u015fekilde yak\u0131nsamas\u0131n\u0131 sa\u011flar. Bu teknik, Transformer tabanl\u0131 modellerde yayg\u0131n olarak kullan\u0131lan bir yak\u0131nsama h\u0131zland\u0131rma stratejisidir.
\n * Daha Az Katman veya Hafif Dikkat Mekanizmalar\u0131:<\/strong> Performans gereksinimlerine ba\u011fl\u0131 olarak, kod \u00e7\u00f6z\u00fcc\u00fc katman say\u0131s\u0131 optimize edilebilir veya daha hafif dikkat mekanizmalar\u0131 (\u00f6rne\u011fin, seyreltilmi\u015f dikkat – sparse attention) kullan\u0131labilir.<\/p>\nRT-DETR’\u0131n Temel Yenilikleri (\u00d6zet)<\/h4>\n
\n* Hibrit Kodlay\u0131c\u0131:<\/strong> \u00c7ok \u00f6l\u00e7ekli \u00f6zellik birle\u015ftirmeyi dikkat mekanizmalar\u0131yla g\u00fc\u00e7lendirerek ve sorgu se\u00e7imi ile daha iyi ba\u015flang\u0131\u00e7 sorgular\u0131 \u00fcreterek kodlay\u0131c\u0131n\u0131n verimlili\u011fini art\u0131r\u0131r.
\n* H\u0131zland\u0131r\u0131lm\u0131\u015f Kod \u00c7\u00f6z\u00fcc\u00fc:<\/strong> Anlaml\u0131 sorgular ve yard\u0131mc\u0131 tahmin ba\u015fl\u0131klar\u0131 sayesinde kod \u00e7\u00f6z\u00fcc\u00fcn\u00fcn yak\u0131nsamas\u0131n\u0131 ve tahmin yetene\u011fini geli\u015ftirir.
\n* Dinamik Sorgu Mekanizmalar\u0131:<\/strong> Statik sorgular yerine, g\u00f6r\u00fcnt\u00fc i\u00e7eri\u011fine duyarl\u0131 sorgu \u00fcretimi, modelin daha h\u0131zl\u0131 ve do\u011fru tahminler yapmas\u0131na olanak tan\u0131r.<\/p>\nUygulama Detaylar\u0131 ve Pratik Yakla\u015f\u0131mlar<\/h3>\n
Geli\u015ftirme Ortam\u0131 ve K\u00fct\u00fcphaneler<\/h4>\n
\n* Derin \u00d6\u011frenme \u00c7er\u00e7evesi:<\/strong> Baidu, kendi PaddlePaddle \u00e7er\u00e7evesini kullan\u0131yor olsa da, RT-DETR genellikle PyTorch implementasyonlar\u0131 arac\u0131l\u0131\u011f\u0131yla da eri\u015filebilirdir. PyTorch, esnekli\u011fi ve geni\u015f topluluk deste\u011fi nedeniyle tercih edilebilir.
\n* Veri \u0130\u015fleme ve G\u00f6rselle\u015ftirme:<\/strong> NumPy ve OpenCV k\u00fct\u00fcphaneleri, g\u00f6r\u00fcnt\u00fc i\u015fleme, veri manip\u00fclasyonu ve sonu\u00e7lar\u0131n g\u00f6rselle\u015ftirilmesi i\u00e7in vazge\u00e7ilmezdir.
\n* H\u0131zland\u0131rma:<\/strong> Modelin e\u011fitim ve \u00e7\u0131kar\u0131m a\u015famalar\u0131nda y\u00fcksek performans i\u00e7in NVIDIA CUDA ve cuDNN k\u00fct\u00fcphaneleri ile uyumlu bir GPU gereklidir.<\/p>\nVeri Seti Haz\u0131rl\u0131\u011f\u0131<\/h4>\n
\n* Etiketleme:<\/strong> Veri setindeki her g\u00f6r\u00fcnt\u00fc i\u00e7in nesnelerin s\u0131n\u0131rlay\u0131c\u0131 kutular\u0131 ve s\u0131n\u0131f etiketleri do\u011fru bir \u015fekilde etiketlenmelidir. COCO format\u0131 (JSON tabanl\u0131) genellikle Transformer tabanl\u0131 modeller i\u00e7in tercih edilen bir formatt\u0131r.
\n* G\u00f6r\u00fcnt\u00fc \u00d6n \u0130\u015fleme:<\/strong>
\n * Yeniden Boyutland\u0131rma (Resizing):<\/strong> Modeller genellikle sabit boyutlu girdiler bekler (\u00f6rne\u011fin 640×640 veya 800×800). G\u00f6r\u00fcnt\u00fcler bu boyutlara \u00f6l\u00e7eklendirilirken, en boy oran\u0131n\u0131n korunmas\u0131 (padding ile) veya do\u011frudan boyutland\u0131rma gibi stratejiler kullan\u0131labilir.
\n * Normalizasyon:<\/strong> Piksel de\u011ferleri genellikle [0, 1] aral\u0131\u011f\u0131na veya belirli bir ortalama ve standart sapma de\u011ferlerine g\u00f6re normalle\u015ftirilir.
\n * Veri Art\u0131rma (Data Augmentation):<\/strong> Modelin genelleme yetene\u011fini art\u0131rmak i\u00e7in d\u00f6nd\u00fcrme, \u00e7evirme, parlakl\u0131k\/kontrast ayar\u0131, mozaikleme (mosaic) gibi teknikler uygulan\u0131r. \u00d6zellikle Transformer tabanl\u0131 modellerde veri art\u0131rma, yak\u0131nsama h\u0131z\u0131n\u0131 ve nihai performans\u0131 \u00f6nemli \u00f6l\u00e7\u00fcde etkileyebilir.<\/p>\nModelin E\u011fitimi<\/h4>\n
\n* Hiperparametre Ayarlar\u0131:<\/strong>
\n * \u00d6\u011frenme Oran\u0131 (Learning Rate):<\/strong> Genellikle k\u00fc\u00e7\u00fck bir ba\u015flang\u0131\u00e7 \u00f6\u011frenme oran\u0131 (\u00f6rne\u011fin 1e-4 veya 1e-5) ile ba\u015flan\u0131r ve e\u011fitim ilerledik\u00e7e azalt\u0131l\u0131r (scheduler).
\n * Batch Boyutu (Batch Size):<\/strong> GPU belle\u011fine ve veri setinin b\u00fcy\u00fckl\u00fc\u011f\u00fcne ba\u011fl\u0131 olarak ayarlan\u0131r. Daha b\u00fcy\u00fck batch boyutlar\u0131, daha kararl\u0131 gradyanlar sa\u011flayabilir ancak daha fazla bellek gerektirir.
\n * Optimizat\u00f6r:<\/strong> AdamW, Transformer tabanl\u0131 modeller i\u00e7in yayg\u0131n olarak kullan\u0131lan ve iyi performans g\u00f6steren bir optimizat\u00f6rd\u00fcr.
\n * Epok Say\u0131s\u0131:<\/strong> DETR modelleri genellikle uzun e\u011fitim s\u00fcreleri gerektirse de, RT-DETR’\u0131n optimizasyonlar\u0131 sayesinde daha az epokta y\u00fcksek performansa ula\u015f\u0131labilir.
\n* Kay\u0131p Fonksiyonlar\u0131 (Loss Functions):<\/strong>
\n * S\u0131n\u0131fland\u0131rma Kayb\u0131:<\/strong> Focal Loss veya Cross-Entropy Loss, s\u0131n\u0131fland\u0131rma g\u00f6revleri i\u00e7in kullan\u0131l\u0131r.
\n * S\u0131n\u0131rlay\u0131c\u0131 Kutu Kayb\u0131:<\/strong> L1 Loss (kutunun koordinatlar\u0131 i\u00e7in) ve GIoU\/DIoU\/CIoU Loss (kutu \u00f6rt\u00fc\u015fme oran\u0131 i\u00e7in) kombinasyonu kullan\u0131l\u0131r. \u00d6zellikle IoU tabanl\u0131 kay\u0131plar, kutu regresyonunun kalitesini art\u0131r\u0131r.
\n* E\u011fitim S\u00fcreci ve \u0130zleme:<\/strong> E\u011fitimin ilerlemesini izlemek i\u00e7in TensorBoard gibi ara\u00e7lar kullan\u0131l\u0131r. Bu, kay\u0131p de\u011ferlerini, do\u011fruluk metriklerini ve \u00f6\u011frenme oran\u0131n\u0131 g\u00f6rselle\u015ftirmeye yard\u0131mc\u0131 olur. Model, do\u011frulama veri setindeki performans\u0131na g\u00f6re kaydedilir.<\/p>\nGer\u00e7ek Zamanl\u0131 \u00c7\u0131kar\u0131m (Inference)<\/h4>\n
\n* Girdi Alma:<\/strong> G\u00f6r\u00fcnt\u00fc veya video ak\u0131\u015f\u0131ndan (kamera, dosya, a\u011f ak\u0131\u015f\u0131) girdi al\u0131n\u0131r.
\n* \u00d6n \u0130\u015fleme:<\/strong> Girdi g\u00f6r\u00fcnt\u00fcleri, e\u011fitim s\u0131ras\u0131nda kullan\u0131lan ayn\u0131 \u00f6n i\u015fleme ad\u0131mlar\u0131ndan (yeniden boyutland\u0131rma, normalizasyon) ge\u00e7irilir.
\n* Modelden Tahmin Alma:<\/strong> \u00d6n i\u015flenmi\u015f g\u00f6r\u00fcnt\u00fc modelinize beslenir ve model, s\u0131n\u0131rlay\u0131c\u0131 kutu koordinatlar\u0131, g\u00fcven skorlar\u0131 ve s\u0131n\u0131f etiketleri \u015feklinde tahminler \u00fcretir.
\n* Sonu\u00e7lar\u0131n \u0130\u015flenmesi ve G\u00f6rselle\u015ftirilmesi:<\/strong>
\n * Tahmin edilen kutu koordinatlar\u0131, orijinal g\u00f6r\u00fcnt\u00fc boyutlar\u0131na geri \u00f6l\u00e7eklendirilir.
\n * D\u00fc\u015f\u00fck g\u00fcven skoruna sahip tahminler filtrelenir.
\n * Sonu\u00e7lar, g\u00f6r\u00fcnt\u00fcn\u00fcn \u00fczerine s\u0131n\u0131rlay\u0131c\u0131 kutular \u00e7izilerek ve s\u0131n\u0131f etiketleri eklenerek g\u00f6rselle\u015ftirilir.
\n* Performans \u00d6l\u00e7\u00fctleri:<\/strong> Ger\u00e7ek zamanl\u0131 performans, FPS (Frame Per Second – saniyedeki kare say\u0131s\u0131) ile \u00f6l\u00e7\u00fcl\u00fcr. Do\u011fruluk ise mAP (mean Average Precision) gibi metriklerle de\u011ferlendirilir.<\/p>\nOptimizasyon ve Da\u011f\u0131t\u0131m<\/h4>\n
\n* ONNX D\u0131\u015fa Aktar\u0131m\u0131 ve TensorRT ile H\u0131zland\u0131rma:<\/strong> Model, ONNX (Open Neural Network Exchange) format\u0131na d\u00f6n\u00fc\u015ft\u00fcr\u00fclebilir. NVIDIA GPU’lar i\u00e7in TensorRT, ONNX modellerini daha da optimize ederek \u00e7\u0131kar\u0131m h\u0131z\u0131nda \u00f6nemli art\u0131\u015flar sa\u011flar.
\n* Donan\u0131m H\u0131zland\u0131rma:<\/strong> GPU’lar, NPU’lar (Neural Processing Units) veya di\u011fer \u00f6zel yapay zeka h\u0131zland\u0131r\u0131c\u0131lar, modelin \u00e7\u0131kar\u0131m performans\u0131n\u0131 art\u0131rmak i\u00e7in kullan\u0131l\u0131r.
\n* Mobil veya G\u00f6m\u00fcl\u00fc Sistemlere Da\u011f\u0131t\u0131m:<\/strong> Mobil veya IoT cihazlar gibi kaynak k\u0131s\u0131tl\u0131 ortamlarda da\u011f\u0131t\u0131m i\u00e7in model s\u0131k\u0131\u015ft\u0131rma teknikleri (pruning, distillation) veya daha k\u00fc\u00e7\u00fck RT-DETR varyantlar\u0131 tercih edilebilir.<\/p>\nPerformans De\u011ferlendirmesi ve Kar\u015f\u0131la\u015ft\u0131rmalar<\/h3>\n
\n* Faster R-CNN ve DETR ile Kar\u015f\u0131la\u015ft\u0131rma:<\/strong> Faster R-CNN gibi iki a\u015famal\u0131 alg\u0131lay\u0131c\u0131lar y\u00fcksek do\u011fruluk sunsa da, RT-DETR bunlardan kat kat daha h\u0131zl\u0131d\u0131r. DETR’\u0131n orijinal versiyonuna k\u0131yasla ise RT-DETR, ayn\u0131 veya daha y\u00fcksek do\u011fruluk seviyelerinde \u00e7ok daha h\u0131zl\u0131 \u00e7\u0131kar\u0131m yapar ve \u00e7ok daha h\u0131zl\u0131 yak\u0131nsar. Bu, RT-DETR’\u0131n hibrit kodlay\u0131c\u0131 ve sorgu se\u00e7imi gibi yeniliklerinin do\u011frudan bir sonucudur.
\n* K\u00fc\u00e7\u00fck Nesne Alg\u0131lama:<\/strong> Geleneksel DETR modelleri k\u00fc\u00e7\u00fck nesneleri alg\u0131lamada zorlan\u0131rken, RT-DETR’\u0131n \u00e7ok \u00f6l\u00e7ekli \u00f6zellik birle\u015ftirme ve daha ak\u0131ll\u0131 sorgu mekanizmalar\u0131 sayesinde bu alandaki performans\u0131 \u00f6nemli \u00f6l\u00e7\u00fcde iyile\u015fmi\u015ftir. Ancak, k\u00fc\u00e7\u00fck nesnelerin yo\u011fun oldu\u011fu senaryolarda hala optimizasyon potansiyeli bulunmaktad\u0131r.
\n* Donan\u0131m Ba\u011f\u0131ml\u0131l\u0131\u011f\u0131:<\/strong> RT-DETR’\u0131n performans\u0131, kullan\u0131lan donan\u0131ma g\u00f6re de\u011fi\u015fiklik g\u00f6sterir. G\u00fc\u00e7l\u00fc GPU’lar (\u00f6rne\u011fin NVIDIA A100, V100) y\u00fcksek FPS de\u011ferleri sunarken, daha d\u00fc\u015f\u00fck seviyeli GPU’larda (\u00f6rne\u011fin T4, 2080Ti) bile hala ger\u00e7ek zamanl\u0131 performans elde edilebilir. Nicemleme ve TensorRT gibi optimizasyonlar, daha k\u0131s\u0131tl\u0131 donan\u0131mlarda bile kabul edilebilir h\u0131zlara ula\u015fmay\u0131 m\u00fcmk\u00fcn k\u0131lar.<\/p>\nZorluklar ve Gelecek Y\u00f6nelimler<\/h3>\n
Kar\u015f\u0131la\u015f\u0131labilecek Zorluklar<\/h4>\n
\n* Modelin Karma\u015f\u0131kl\u0131\u011f\u0131 ve Hiperparametre Ayarlar\u0131:<\/strong> RT-DETR’\u0131n mimarisi, geleneksel tek a\u015famal\u0131 alg\u0131lay\u0131c\u0131lara g\u00f6re daha karma\u015f\u0131kt\u0131r. Bu, optimal hiperparametre ayarlar\u0131n\u0131 bulmay\u0131 ve modeli belirli bir veri setine uyarlamay\u0131 zorla\u015ft\u0131rabilir.
\n* K\u00fc\u00e7\u00fck Nesnelerde Alg\u0131lama Zorluklar\u0131:<\/strong> \u00c7ok \u00f6l\u00e7ekli \u00f6zellik birle\u015ftirme ile iyile\u015fme olsa da, a\u015f\u0131r\u0131 k\u00fc\u00e7\u00fck nesnelerin yo\u011fun oldu\u011fu veya d\u00fc\u015f\u00fck \u00e7\u00f6z\u00fcn\u00fcrl\u00fckl\u00fc g\u00f6r\u00fcnt\u00fclerde RT-DETR’\u0131n performans\u0131 hala iyile\u015ftirme potansiyeli ta\u015f\u0131maktad\u0131r.
\n* Farkl\u0131 Veri Setlerine Genellenebilirlik:<\/strong> Modelin COCO gibi b\u00fcy\u00fck ve \u00e7e\u015fitli veri setlerinde iyi performans g\u00f6stermesi, her \u00f6zel veri setinde ayn\u0131 ba\u015far\u0131y\u0131 garantilemez. Domain adaptasyonu ve daha az etiketli veriyle \u00f6\u011frenme teknikleri \u00f6nem kazanmaktad\u0131r.<\/p>\nGelecek Y\u00f6nelimler<\/h4>\n
\n* S\u0131f\u0131r At\u0131\u015fl\u0131 (Zero-shot) ve Az At\u0131\u015fl\u0131 (Few-shot) \u00d6\u011frenme Yetenekleri:<\/strong> RT-DETR’\u0131n k\u00fcresel ba\u011flam\u0131 anlama yetene\u011fi, onu yeni veya az say\u0131da \u00f6rnekle kar\u015f\u0131la\u015f\u0131lan nesneleri tan\u0131ma konusunda daha yetenekli hale getirebilir. Bu alandaki ara\u015ft\u0131rmalar, modelin genelleme yetene\u011fini art\u0131racakt\u0131r.
\n* Multimodal Alg\u0131lama:<\/strong> Sadece g\u00f6rsel verilerle de\u011fil, metin, ses veya di\u011fer sens\u00f6r verileriyle birlikte \u00e7al\u0131\u015fan multimodal RT-DETR modelleri, daha zengin ve ba\u011flama duyarl\u0131 alg\u0131lama yetenekleri sunabilir.
\n* Edge Cihazlarda Daha \u0130yi Performans:<\/strong> Model nicemleme, TensorRT entegrasyonu ve \u00f6zel donan\u0131m h\u0131zland\u0131r\u0131c\u0131larla daha iyi entegrasyon, RT-DETR’\u0131n kenar bili\u015fim (edge computing) cihazlar\u0131nda daha yayg\u0131n olarak kullan\u0131lmas\u0131na olanak tan\u0131yacakt\u0131r.
\n* G\u00fcvenlik ve Etik Konular:<\/strong> Yapay zeka modellerinin g\u00fcvenli\u011fi (adversarial attacks) ve etik kullan\u0131m\u0131 (\u00f6nyarg\u0131 tespiti) nesne alg\u0131lama alan\u0131nda da \u00f6nem kazanmaktad\u0131r. RT-DETR gibi modellerin bu konularda nas\u0131l davrand\u0131\u011f\u0131 ve nas\u0131l iyile\u015ftirilebilece\u011fi \u00fczerine ara\u015ft\u0131rmalar devam edecektir.
\n* Video Nesne Alg\u0131lama:<\/strong> RT-DETR’\u0131n ger\u00e7ek zamanl\u0131 yetenekleri, onu video nesne alg\u0131lama (object tracking) ve eylem tan\u0131ma gibi g\u00f6revler i\u00e7in ideal bir aday haline getirmektedir. Zamansal bilgiyi entegre eden RT-DETR tabanl\u0131 modeller, bu alanlarda yeni ufuklar a\u00e7abilir.<\/p>\nSonu\u00e7<\/h3>\n