thu-mig/yolov10

★ 11,297PythonAudience · researcherComplexity · 4/5Setup · moderate

Mindmap

mindmap
  root((YOLOv10))
    What it does
      Object detection
      Bounding boxes
      Real-time speed
    Tech stack
      Python
      PyTorch
      Hugging Face
    Model sizes
      Compact edge models
      High accuracy models
    Use cases
      Security cameras
      Photo analysis
      Custom training

mindmap root((YOLOv10)) What it does Object detection Bounding boxes Real-time speed Tech stack Python PyTorch Hugging Face Model sizes Compact edge models High accuracy models Use cases Security cameras Photo analysis Custom training

Click or tap to explore — scroll the page freely

Things people build with this

USE CASE 1

Run real-time object detection on security camera footage to spot and label people, vehicles, or other objects.

USE CASE 2

Add bounding box detection to a Python app that processes product photos or medical images.

USE CASE 3

Fine-tune the model on a custom image dataset to recognize objects specific to your industry.

USE CASE 4

Deploy the compact model variant on edge hardware with limited computing power for on-device inference.

Tech stack

PythonPyTorchHugging Face

Getting it running

Difficulty · moderate Time to first run · 30min

Requires a GPU for practical real-time inference speed, pip install works but CUDA is needed for meaningful performance.

In plain English

YOLOv10 is a computer vision model that identifies and locates objects within images and video. You show it a picture and it draws bounding boxes around things it recognizes, such as cars, people, or animals, along with a label for each. It is part of the YOLO family of models, which are built around speed and can run in real time. This version was developed by researchers at Tsinghua University and published at NeurIPS 2024. The main technical contribution is removing a post-processing step called NMS (non-maximum suppression) that previous YOLO versions required after generating predictions. NMS is a cleanup pass that filters out duplicate detections, but it adds latency and complicates deployment on certain hardware. YOLOv10 is trained to avoid producing duplicates in the first place, so no NMS step is needed when running the model. The paper reports that this makes it 1.8 times faster than a comparable competing model at similar accuracy. Several model sizes are provided, from a compact version suited for devices with limited computing power to larger versions aimed at higher accuracy. Pre-trained checkpoints are available on Hugging Face. The model can be tested through a browser-based demo, run in a Google Colab notebook, or installed as a Python package for integration into custom projects. The README also promotes a follow-up project called YOLOE, which extends this work to open-vocabulary detection: recognizing objects that go beyond a fixed predefined list of categories, by accepting text or visual prompts at inference time. This is a research codebase intended for practitioners familiar with Python and machine learning workflows.

Copy-paste prompts

Prompt 1

Using YOLOv10, write Python code to run inference on a folder of images and save the results with bounding boxes drawn.

Prompt 2

How do I load a YOLOv10 checkpoint from Hugging Face and run it on a live webcam stream in Python?

Prompt 3

I want to fine-tune YOLOv10 on my own labeled dataset. Walk me through the training config and required data format.

Prompt 4

When should I pick the compact YOLOv10 model variant versus the larger high-accuracy version?

Prompt 5

How does YOLOv10 avoid needing NMS post-processing, and what does that mean for deployment on embedded devices?

Open on GitHub → Explain another repo

← thu-mig on gitmyhub — every repo by this author, as a profile.

Verify against the repo before relying on details.