FastYOLO

FastYOLO is a high-performance video object detection library powered by Ultralytics YOLO v11. It is designed to be fast, efficient, and easy to use. FastYOLO maximizes GPU utilization through parallel video decoding and model inference:

1. Video decoding is performed asynchronously on CPU using FFmpeg
2. Model inference runs concurrently on GPU using PyTorch
3. A prefetch queue maintains a buffer of decoded frames ready for inference

This architecture eliminates the typical bottleneck where GPU processing waits for CPU video decoding, resulting in significantly higher throughput compared to sequential processing.

Installation

First make sure ffmpeg is installed.

pip install fastyolo

Usage

import torch
from ultralytics import YOLO
from fastyolo import Video

# Initialize model as usual with Ultralytics
device = "cuda" if torch.cuda.is_available() else "cpu"
model = YOLO("yolo11n.pt", task="detect").to(device)
dtype = next(model.parameters()).dtype

# Create video reader
video = Video(
    "traffic.mp4",
    # 640x640 is the default resolution for YOLO v11
    width=640,
    height=640,
    device=device,
    dtype=dtype
)

# Process video frames in batches
for batch in video:
    # Run inference
    results = model.predict(batch)
    
    # Process results
    for result in results:
        boxes = result.boxes 

Todo

• Optimize Yolo models including post processing
• Add support for other tasks
• Add GPU-accelerated video decoding using torchcodec and other backends
• Add comprehensive benchmarks