DOCSEARCHbyAlgolia
Submit Docs

PaddleX Documentation

repository·release/3.7·Indexed 27 days ago

https://github.com/paddlepaddle/paddlex

A low-code AI development tool built on PaddlePaddle providing end-to-end workflows for training and deploying pre-trained models. It features 33 model pipelines and 39 functional modules across domains such as OCR, object detection, time-series analysis, and computer vision. Supports high-performance inference, service-based deployment via Triton Inference Server, and on-device deployment across various hardware including NVIDIA GPU, Ascend, Kunlunxin, and Cambricon.

Tokens
170.1K
Snippets
434
Records
894
Agent score
91%

What's inside PaddleX

  2. Overview of Rotated Object Detection Module

    release/3.7
    The Rotated Object Detection module is designed to detect objects using Rotated Bounding Boxes (RBB). Unlike standard horizontal bounding boxes, rotated boxes include angle information, allowing the width and height of the rectangle to be non-parallel to the image coordinate axes. This approach reduces background noise within the bounding box and is particularly useful for scenarios like remote sensing.

  3. Overview of Layout Analysis Module

    release/3.7
    The Layout Analysis module performs layout area detection, instance segmentation, and reading order prediction. It identifies various document elements (e.g., text, charts, images, formulas, titles) and provides their bounding boxes, precise contour masks, and reading order indices. This is useful for document understanding and information extraction workflows.

  4. Overview of 3D Multimodal Fusion Detection Module

    release/3.7
    The 3D multimodal fusion detection module is designed for computer vision and autonomous driving tasks. It locates and marks 3D coordinates and detection box information for specific targets in images or videos. The module typically outputs 3D bounding boxes, which serve as inputs for subsequent target recognition modules.

  5. Overview of Document Image Orientation Classification Module

    release/3.7
    The Document Image Orientation Classification module is designed to identify and correct the orientation of document images (such as scanned documents or ID cards). This module is intended to be used as a pre-processing step in OCR pipelines to ensure text regions are correctly oriented, thereby improving OCR accuracy. It classifies images into four categories: 0°, 90°, 180°, and 270°.

  6. Overview of Image Multi-Label Classification Module

    release/3.7
    The image multi-label classification module assigns multiple relevant categories to an input image simultaneously. Unlike standard classification which provides a single label, this module identifies all applicable predefined categories (e.g., labeling an image as both 'cat' and 'dog'). It is a core component for computer vision systems requiring complex scene understanding.

  7. Overview of Human Detection Module

    release/3.7
    The Human Detection module is a computer vision subtask of object detection. It identifies the presence of pedestrians in images or videos and provides specific location information (bounding boxes) for each detected person. It is designed for applications like intelligent video surveillance, human behavior analysis, autonomous driving, and robotics.

  8. Overview of PaddleX 3.0 Deployment Methods

    release/3.7

    PaddleX 3.0 provides three primary deployment strategies depending on your performance and environment requirements:

    • High-Performance Inference: Uses a specialized plugin to optimize model inference and pre/post-processing for maximum end-to-end speed. Best for production environments with strict latency requirements.
    • Serving Deployment: Encapsulates inference functionality as network services, allowing clients to access models via network requests. Best for low-cost service-oriented architectures.
    • On-Device Deployment: Deploys models directly onto edge devices (e.g., Android) to enable local data processing without remote server reliance.

  9. Overview of Seal Text Detection Module

    release/3.7
    The Seal Text Detection module is a subtask of OCR (Optical Character Recognition) designed to locate and mark regions containing seal text within an image. It outputs multi-point bounding boxes around text regions. These bounding boxes are intended to be used as inputs for subsequent distortion correction and text recognition modules to identify the actual textual content of the seal. This module is critical for document processing tasks like contract comparison, inventory auditing, and invoice verification.

  10. Overview of Text Line Orientation Classification Module

    release/3.7
    The text line orientation classification module identifies and corrects the orientation of text lines. This is particularly useful in document scanning or photography of licenses and certificates where the capture device might be rotated. By using image classification to determine if a text line is at 0 or 180 degrees, the module can apply post-processing to rotate the image, ensuring standard OCR pipelines can process the data accurately.

  11. Overview of Vehicle Attribute Recognition Module

    release/3.7
    The Vehicle Attribute Recognition module identifies and labels specific vehicle attributes (e.g., vehicle type, color, license plate number) within images or videos. It typically outputs bounding boxes containing attribute information, which can be used as input for downstream tasks like vehicle tracking or vehicle re-identification.

  12. Overview of Time Series Anomaly Detection Module

    release/3.7
    The Time Series Anomaly Detection module in PaddleX is designed to identify abnormal points or periods in time series data that deviate from expected patterns, trends, or periodic regularities. It is useful for detecting system failures, external shocks, data entry errors, or rare events to support risk assessment and business decision-making.