Home Blog Forums Docs Downloads Training Join

Topics Platforms and Tools Industries Resources

Explore Computer Vision SDKs and Libraries

Get the flexibility and reliability you need to integrate powerful visual perception into your application.

Whether you're developing an autonomous vehicle's driver assistance system or a sophisticated industrial system, your
computer vision pipeline needs to be versatile. This means supporting deployment from the cloud to the edge, while
remaining stable and production-ready. NVIDIA SDKs and libraries deliver the right solution for your unique needs.

Benefits of NVIDIA Computer Vision Software

High Accuracy High Throughput Versatile Deployment Flexible Interoperability

Tap into production-ready, pretrained Achieve real-time processing with high- Scale your systems in deployment NVIDIA SDKs are interoperable with each
models and low-code tools for fine throughput SDKs and libraries. environments from the edge to the cloud other and with popular frameworks like
tuning on custom data sets. and data center. TensorFlow and Pytorch.

Computer Vision Pipelines

A computer vision pipeline process begins with decoding the image or video input to make it suitable for analysis. Pre-processing steps follow, where the
data is transformed, normalized, and enhanced to improve the accuracy of subsequent computations. Next is the inference stage, where the data is
analyzed and interpreted to extract meaningful information, such as object detection or image classification results. After inference, post-processing
steps are applied to refine and organize the output. And finally, the processed image or video is encoded to its appropriate format before being output
from the pipeline, ready for further use or presentation.

This complete pipeline, supported by our versatile SDKs and libraries, seamlessly guides you through every step of the computer vision process.

Encoding and Decoding

Achieve 4X faster lossless image decoding and up to 7X higher throughput with video encoding using GPU-accelerated libraries for encoding and
decoding.

Learn more about encoding and decoding SDKs 

Image-Decoding Libraries Hardware-Accelerated Video Encoding Hardware-Accelerated Video Encoding

nvJPEG and nvJPEG 2000 Video Codec SDK Video Processing Framework
(VPF)
C++ libraries for decoding, encoding, and A set of APIs for hardware-accelerated video encode,
transcoding JPEG and JPEG 2000 format images. transcode, and decode on Windows and Linux. Python bindings of C++ libraries from Video Codec
SDK.
Learn More About nvJPEG and nvJPEG 2000  Learn More About Video Codec SDK 
Learn More About VPF 

Pre- and Post-Processing

Move your pre- and post-processing pipelines that are bottlenecked on the CPU to the GPU and achieve up to a 49X end-to-end speedup.

Learn more about pre- and post-processing SDKs 

Open Source Pre- and Post-Processing Library Portable Data Processing Pipeline Embedded Computer Vision and Image Processing Library

CV-CUDA Data Loading Library (DALI) Vision Programming Interface (VPI)

An open-source, low-level library that easily Load and pre-process image, audio, and video data A low-level library for running highly optimized
integrates into existing custom CV applications to using GPUs. DALI can be used directly in TensorFlow, algorithms for processing pipelines on embedded
accelerate video and image processing. PyTorch, MXNet, and PaddlePaddle models. devices. Efficiently distribute workloads across
multiple compute engines with support for both C
Learn More About CV-CUDA  Learn More About DALI  and Python.

Learn More About VPI 

Image and Signal-Processing Library Motion Flow Generation

NVIDIA Performance Primitives (NPP) Optical Flow SDK

A comprehensive set of image, video, and signal Detect and track objects in successive video frames,
processing functions for high-performance, low-level interpolate, or extrapolate video frames to improve
image or signal processing of large images (e.g., smoothness of video playback and compute flow
20kx20k pixels). vectors.

Learn More About NPP  Learn More About Optical Flow SDK 

Training

Simplify the model training process and create and train customized AI models in a fraction of the time as before with GPU-accelerated transfer learning
and synthetic data.

Learn more about training SDKs 

Synthetic Data Generation Open-Source Training Toolkit

NVIDIA Omniverse™ Replicator NVIDIA TAO Toolkit

Fine-tune pretrained models with custom, physically Create highly accurate, customized, and enterprise-
accurate 3D synthetic visual data generated in ready AI models with this low-code toolkit and
minutes or hours using OpenUSD rather than deploy them on any device, at the edge or the cloud.
months.
Learn More About TAO Toolkit 
Learn More About Omniverse Replicator 

Inference

Optimize and speed up inference performance by 36X over CPU-only platforms during inference.

Learn more about inference SDKs 

Inference Optimizer and Runtime Open Source Inference Serving Software Streaming Analytics Toolkit

NVIDIA TensorRT™ NVIDIA Triton™ NVIDIA DeepStream SDK

A high-performance deep learning inference SDK Streamlines and standardizes AI inference by AI-based multi-sensor processing, video, audio, and
that delivers low latency and high throughput for enabling teams to deploy, run, and scale trained ML image understanding to enable real-time analytics.
inference applications. or DL models from any framework on any GPU- or
CPU-based infrastructure. Learn More About DeepStream SDK 
Learn More About TensorRT 
Learn More About Triton 

The world of computer vision solutions is powered by NVIDIA.

Join Us

Follow NVIDIA Developer

Sign up for NVIDIA News Subscribe
Find more news and tutorials on NVIDIA Technical Blog

Privacy Policy | Your Privacy Choices | Terms of Use | Accessibility | Corporate Policies | Contact

Copyright © 2026 NVIDIA Corporation