Image Streaming API with YOLO

An object detection system using YOLOv26n for benchmarking WebSocket-based image streaming latency between Python and Go server implementations.

This repository was developed with assistance from Claude Code and Gemini.

Overview

This project compares the end-to-end latency of real-time image streaming using WebSockets between:

• Python server — built with FastAPI and websockets
• Go server — built with the standard net/http and gorilla/websocket

Both servers run the same YOLO inference pipeline and provide an identical WebSocket interface.

Getting Started

1. Download Model Weights

$ cd py_server
$ uv run utils/get_model.py

2. Start the Servers

$ docker compose build
$ docker compose up

• Go server: http://localhost:8001
• Python server: http://localhost:8000

3-1. Run the Test

$ cd py_server
$ uv run utils/test.py \
    --video-in "assets/test.mp4" \
    --bench-out "results/go_result.csv" \
    --server-type "go" \
    --clients 6

--server-type must be set to py or go.

3-2. Run the Client

$ cd client
$ flutter pub get
$ flutter run

Test Results

• OS: macOS 26.2
• CPU: Apple M4
• Python: 3.12.12
• Go: 1.22.8

Latency

Scenario: 20 concurrent clients

Metric	Go Server	Python Server
Average Latency	413.09 ms	437.10 ms
Average FPS	2.53 ms	2.35 ms
P95 Latency	542.26 ms	549.43 ms
P99 Latency	613.48 ms	623.11 ms

CPU Usage

Scenario: 6 concurrent clients