Which model to use for vision

• YoloV5 ← current
• YoloV8
  • https://docs.ultralytics.com/guides/object-counting/#what-is-object-counting
  • https://docs.ultralytics.com/guides/nvidia-jetson/#nvidia-jetson-orin-yolov8-benchmarks
• YoloV6-v3 at 1280px, GPL-3.0, from source code
• https://github.com/lyuwenyu/RT-DETR with 54.8 mAP
• https://github.com/jozhang97/DETA 63.5 mAP
• MCUNet ? https://github.com/mit-han-lab/mcunet

Zero-shot models seem pretty bad:

• YoloWorld
• OWL + CLIP
• OWL + ViT
• LLaVA 1.5 = detect everything
• NanoOWL github.com/NVIDIA-AI-IOT/nanoowl

Hardware

• RTX
• A100

Runtimes

• Pytorch
• TensorRT
• ONNX