YOLOv10 – Real-Time End-to-End Object Detection

YOLOv10 is a real-time object detection model introduced in 2024 by Ao Wang, Hui Chen, Lihao Liu, Kai Chen, Zijia Lin, Jungong Han, and Guiguang Ding. The model was presented in the paper YOLOv10: Real-Time End-to-End Object Detection and was accepted at NeurIPS 2024.

YOLOv10 focuses on improving both detection efficiency and accuracy while removing one of the traditional post-processing steps used by many object detectors: Non-Maximum Suppression, commonly known as NMS.

The main goals of YOLOv10 are:

End-to-end object detection
NMS-free inference
Lower latency
Reduced computational redundancy
Strong speed-to-accuracy balance
Efficient model scaling
Better deployment performance

What Is YOLOv10?

YOLOv10 is a single-stage real-time object detector designed to produce final detections directly from a neural network.

Traditional YOLO-style detectors often generate many overlapping bounding boxes and then use Non-Maximum Suppression to remove duplicates.

YOLOv10 is designed to avoid this additional post-processing step.

The model uses a training strategy called consistent dual assignments to enable NMS-free detection while maintaining competitive detection performance.

YOLOv10 End-to-End Detection

One of the most important characteristics of YOLOv10 is its end-to-end detection design.

A traditional detection pipeline may look like:

Image → Neural Network → Bounding Boxes → NMS → Final Detections

YOLOv10 aims to simplify this to:

Image → Neural Network → Final Detections

Removing NMS can reduce processing overhead and make the detection pipeline easier to deploy.

The YOLOv10 researchers specifically identified NMS as a source of inference latency and a limitation for fully end-to-end YOLO deployment.

What Is Non-Maximum Suppression?

Non-Maximum Suppression is a post-processing technique commonly used in object detection.

A detector may predict several overlapping bounding boxes around the same object.

NMS generally:

Sorts detections by confidence.
Keeps the strongest prediction.
Compares overlapping boxes.
Removes lower-confidence duplicate detections.

Although effective, this adds another processing stage after the neural network.

YOLOv10 was designed to remove this dependency.

NMS-Free Training

YOLOv10 introduces consistent dual assignments for NMS-free training.

This is one of the main innovations of the model.

The goal is to train the detector so that it can generate high-quality final predictions directly, without relying on NMS during inference.

This can provide benefits such as:

Lower inference latency
Simpler deployment pipelines
Reduced post-processing
Better end-to-end optimization

Consistent Dual Assignments

YOLOv10 uses a dual label assignment strategy during training.

The general idea is to combine different assignment behaviors in a consistent way so that the network can learn effectively while still producing one-to-one style predictions suitable for end-to-end detection.

This strategy enables YOLOv10 to preserve strong detection performance while removing the need for NMS at inference time.

Efficiency-Accuracy Driven Design

YOLOv10 does not focus only on removing NMS.

The researchers also performed a broad analysis of different YOLO model components and identified areas containing unnecessary computational cost.

They introduced a holistic efficiency-accuracy driven model design strategy.

This means different components were evaluated according to both:

Detection accuracy
Computational efficiency

The goal was to remove unnecessary complexity while preserving or improving detection quality.

Reducing Computational Redundancy

YOLOv10 attempts to reduce computational redundancy throughout the network.

Instead of increasing model complexity without careful analysis, the architecture is designed to use computational resources more efficiently.

The researchers optimized multiple components from both efficiency and accuracy perspectives.

According to the original repository, these optimizations reduce computational overhead while improving model capability.

YOLOv10 Model Variants

YOLOv10 is available in multiple sizes.

The main variants are:

YOLOv10-N
YOLOv10-S
YOLOv10-M
YOLOv10-B
YOLOv10-L
YOLOv10-X

Each model targets a different balance between:

Accuracy
Parameters
FLOPs
Latency
Hardware requirements

The official repository provides pretrained checkpoints for all six variants.

YOLOv10-N

YOLOv10-N is the Nano model.

According to the official COCO results, YOLOv10-N uses approximately:

2.3 million parameters
6.7 GFLOPs
38.5% AP
1.84 ms latency

at a 640-pixel test size under the repository benchmark setup.

YOLOv10-N is intended for lightweight environments where inference speed and low computational cost are important.

YOLOv10-S

YOLOv10-S is the Small variant.

The official repository reports:

7.2 million parameters
21.6 GFLOPs
46.3% AP
2.49 ms latency

at a 640-pixel test size.

This model offers a balance between lightweight deployment and detection accuracy.

YOLOv10-M

YOLOv10-M is the Medium model.

Its official reported results include:

15.4 million parameters
59.1 GFLOPs
51.1% AP
4.74 ms latency

at a 640-pixel input size.

YOLOv10-M is suitable for applications that require higher accuracy while maintaining fast inference.

YOLOv10-B

YOLOv10-B is the Balanced model.

The repository reports approximately:

19.1 million parameters
92.0 GFLOPs
52.5% AP
5.74 ms latency

at a 640-pixel test resolution.

The researchers also compare YOLOv10-B with YOLOv9-C and report that YOLOv10-B achieves similar performance with 46% lower latency and 25% fewer parameters.

YOLOv10-L

YOLOv10-L is the Large model.

The repository reports:

24.4 million parameters
120.3 GFLOPs
53.2% AP
7.28 ms latency

at 640-pixel test size.

This variant provides greater model capacity for applications where accuracy is prioritized over minimum latency.

YOLOv10-X

YOLOv10-X is the largest standard YOLOv10 variant.

The official repository reports:

29.5 million parameters
160.4 GFLOPs
54.4% AP
10.70 ms latency

at 640-pixel resolution.

YOLOv10-X targets high-accuracy environments with stronger GPU resources.

YOLOv10 Performance Table

Model	Parameters	FLOPs	COCO AP	Latency
YOLOv10-N	2.3M	6.7G	38.5%	1.84 ms
YOLOv10-S	7.2M	21.6G	46.3%	2.49 ms
YOLOv10-M	15.4M	59.1G	51.1%	4.74 ms
YOLOv10-B	19.1M	92.0G	52.5%	5.74 ms
YOLOv10-L	24.4M	120.3G	53.2%	7.28 ms
YOLOv10-X	29.5M	160.4G	54.4%	10.70 ms

These values are reported by the official YOLOv10 repository using a test size of 640 pixels.

YOLOv10 vs RT-DETR

The official YOLOv10 results compare the model with RT-DETR.

According to the authors, YOLOv10-S is approximately 1.8× faster than RT-DETR-R18 at a similar AP level on COCO.

YOLOv10-S also uses approximately 2.8× fewer parameters and FLOPs in that comparison.

This illustrates YOLOv10’s focus on improving the speed-accuracy tradeoff.

YOLOv10 vs YOLOv9

YOLOv10 was also compared directly with YOLOv9.

The official repository reports that:

YOLOv10-B has 46% lower latency and 25% fewer parameters than YOLOv9-C at similar performance.

The main difference is that YOLOv10 places greater emphasis on:

End-to-end detection
NMS-free inference
Reducing computational redundancy
Optimizing latency and accuracy together

YOLOv10 and COCO

YOLOv10 was evaluated on the MS COCO object detection benchmark.

The repository reports AP values for all standard YOLOv10 models using a 640-pixel test resolution.

COCO is widely used for evaluating object detectors because it contains many object categories and challenging image conditions.

YOLOv10 Training

The official repository includes training support through the YOLOv10 implementation.

A typical training configuration uses:

COCO dataset
640-pixel images
500 epochs
Configurable batch size
Multi-GPU support

The repository provides both CLI and Python-based training examples.

YOLOv10 Custom Training

YOLOv10 can also be fine-tuned on custom datasets.

A custom object detection dataset generally requires:

Training images
Validation images
Bounding box labels
Object class definitions
Dataset configuration

This makes YOLOv10 suitable for specialized use cases such as:

Vehicle detection
Industrial inspection
Person detection
Wildlife monitoring
Agriculture
Manufacturing
Robotics
Traffic analysis

YOLOv10 Prediction

The repository provides prediction functionality through both CLI and Python workflows.

Models can be loaded using pretrained YOLOv10 checkpoints and used to perform object detection.

The main prediction process includes:

Loading the model.
Loading the input image or media.
Running inference.
Predicting bounding boxes.
Predicting object classes.
Returning final detections.

Because YOLOv10 is designed for NMS-free inference, final predictions can be produced without the traditional NMS post-processing stage.

YOLOv10 Validation

YOLOv10 includes model validation functionality.

The official repository provides commands for evaluating YOLOv10-N, S, M, B, L, and X models on COCO.

Validation helps measure:

Detection accuracy
Average Precision
Model behavior
Performance across object classes

YOLOv10 Export

YOLOv10 supports export for deployment.

The official repository includes examples for:

ONNX
TensorRT

and describes these as end-to-end export workflows.

Exporting makes YOLOv10 easier to deploy in production environments outside the original training setup.

YOLOv10 with ONNX

YOLOv10 models can be exported to ONNX.

ONNX can be useful when deploying models in:

C++ applications
Edge environments
Cross-platform inference systems
OpenCV-based pipelines
Hardware-specific runtimes

The official repository provides commands for end-to-end ONNX export.

YOLOv10 with TensorRT

YOLOv10 can also be exported to TensorRT.

TensorRT is commonly used to optimize neural network inference on NVIDIA GPUs.

The YOLOv10 repository provides examples for exporting TensorRT engines using FP16 precision.

This can help reduce inference latency in deployment environments.

YOLOv10 and OpenVINO

The project also lists community integration with OpenVINO, allowing YOLOv10 models to be used in Intel-oriented deployment environments.

This expands deployment options beyond standard PyTorch inference.

YOLOv10 and Object Tracking

The repository lists integrations with multiple object tracking systems, including:

BoTSORT
DeepOCSORT
OCSORT
HybridSORT
ByteTrack
StrongSORT

through the BoxMOT ecosystem.

This allows YOLOv10 detections to be incorporated into multi-object tracking applications.

YOLOv10 and DeepSORT

YOLOv10 also has community integration with DeepSORT listed in the official repository.

DeepSORT can be combined with object detection to maintain object identities across video frames.

This is useful for:

People tracking
Vehicle tracking
Surveillance
Traffic analysis
Sports analytics

YOLOv10 Demo

The official repository includes a local Gradio demo.

The repository provides an app.py file that can launch a browser-based interface for running YOLOv10 locally.

This makes it easier to test the model without building a custom frontend.

YOLOv10 and PyTorch

The official repository describes itself as the official PyTorch implementation of YOLOv10.

This provides a Python-oriented workflow for:

Training
Validation
Prediction
Export
Fine-tuning

The implementation is also built using components from the Ultralytics codebase and RT-DETR.

YOLOv10 Advantages

End-to-End Detection

YOLOv10 can generate final detections without traditional NMS post-processing.

Lower Latency

Removing NMS and optimizing network components reduces inference latency.

Efficient Model Design

YOLOv10 reduces computational redundancy across multiple parts of the architecture.

Multiple Model Sizes

Users can select Nano, Small, Medium, Balanced, Large, or Extra-Large models.

Strong Accuracy-to-Speed Balance

The model was explicitly designed around balancing detection accuracy and computational cost.

Deployment Support

YOLOv10 supports deployment-oriented formats such as ONNX and TensorRT.

Custom Dataset Training

The model can be trained or fine-tuned for specialized detection tasks.

YOLOv10 Limitations

YOLOv10 also has practical limitations.

Larger Models Need More Hardware

YOLOv10-L and YOLOv10-X require substantially more computation than Nano or Small models.

Accuracy and Speed Tradeoff

Larger models generally improve accuracy but increase latency.

Training Can Be Resource Intensive

Full-scale training on large datasets can require powerful GPUs and significant training time.

Deployment Performance Depends on Export

The repository specifically notes that benchmarking non-exported PyTorch models can produce biased speed results because unnecessary operations may still execute. The authors recommend using exported formats for benchmarking.

YOLOv10 Use Cases

YOLOv10 can be used for many real-time computer vision applications, including:

Person detection
Vehicle detection
Traffic monitoring
Industrial inspection
Robotics
Drone vision
Security systems
Manufacturing
Retail analytics
Agriculture
Wildlife monitoring
Object tracking
Real-time video analysis

Its NMS-free design is particularly useful for applications where inference latency is important.

Why YOLOv10 Is Important

YOLOv10 represents an important evolution in the YOLO family because it targets one of the remaining bottlenecks in real-time object detection: post-processing.

Instead of accepting NMS as a required step, YOLOv10 redesigns the training strategy so the model can produce high-quality end-to-end detections directly.

Its major contributions include:

NMS-free detection
Consistent dual assignments
End-to-end inference
Holistic efficiency-accuracy design
Reduced computational redundancy
Multiple optimized model sizes

These improvements push YOLO toward simpler and faster production deployment.

Frequently Asked Questions

What is YOLOv10?

YOLOv10 is a real-time end-to-end object detection model introduced in 2024.

Who created YOLOv10?

YOLOv10 was introduced by:

Ao Wang
Hui Chen
Lihao Liu
Kai Chen
Zijia Lin
Jungong Han
Guiguang Ding

What is the YOLOv10 paper called?

The paper is titled:

YOLOv10: Real-Time End-to-End Object Detection.

When was YOLOv10 introduced?

YOLOv10 was introduced in 2024 and the implementation is associated with NeurIPS 2024.

What makes YOLOv10 different?

The biggest difference is its focus on NMS-free end-to-end object detection.

Does YOLOv10 require Non-Maximum Suppression?

No. YOLOv10 is specifically designed for NMS-free inference.

What are consistent dual assignments?

Consistent dual assignments are YOLOv10’s training strategy for enabling NMS-free detection while maintaining strong detection accuracy.

What YOLOv10 models are available?

The standard variants are:

YOLOv10-N
YOLOv10-S
YOLOv10-M
YOLOv10-B
YOLOv10-L
YOLOv10-X

What image size is used for YOLOv10 benchmarks?

The official repository reports its standard COCO results using a 640-pixel test size.

Can YOLOv10 be trained on custom data?

Yes. YOLOv10 supports training and fine-tuning on custom object detection datasets.

Does YOLOv10 support ONNX?

Yes. The repository provides end-to-end ONNX export examples.

Does YOLOv10 support TensorRT?

Yes. YOLOv10 provides TensorRT export workflows for optimized NVIDIA GPU inference.

Is YOLOv10 faster than YOLOv9?

The exact answer depends on the models being compared, but the official authors report that YOLOv10-B has 46% lower latency than YOLOv9-C at similar performance.

Is YOLOv10 suitable for real-time applications?

Yes. YOLOv10 was specifically designed for real-time object detection and low-latency deployment.

Is YOLOv10 still useful?

Yes. YOLOv10 is particularly useful for applications that benefit from fast, end-to-end object detection without the additional latency of traditional NMS post-processing.

Conclusion

YOLOv10 advances real-time object detection by redesigning both the training process and network architecture around end-to-end efficiency.

Its most important innovation is the removal of traditional Non-Maximum Suppression through consistent dual assignments, allowing the model to produce final detections directly.

Combined with a holistic efficiency-accuracy driven architecture, reduced computational redundancy, six different model sizes, and deployment support for ONNX and TensorRT, YOLOv10 provides a strong balance between detection accuracy, model size, and inference latency.

This makes YOLOv10 an important step in the evolution of YOLO toward simpler, faster, and more deployment-friendly real-time object detection.