graph LR
Inference_Orchestrator["Inference Orchestrator"]
Class_Name_Loader["Class Name Loader"]
Non_Maximum_Suppression_NMS_["Non-Maximum Suppression (NMS)"]
Bounding_Box_IoU_Calculator["Bounding Box IoU Calculator"]
Coordinate_Format_Converter["Coordinate Format Converter"]
Coordinate_Scaler["Coordinate Scaler"]
Classifier_Application["Classifier Application"]
Inference_Orchestrator -- "Depends on" --> Class_Name_Loader
Inference_Orchestrator -- "Passes outputs to" --> Non_Maximum_Suppression_NMS_
Inference_Orchestrator -- "Utilizes" --> Classifier_Application
Non_Maximum_Suppression_NMS_ -- "Relies on" --> Bounding_Box_IoU_Calculator
Non_Maximum_Suppression_NMS_ -- "Relies on" --> Coordinate_Format_Converter
Classifier_Application -- "Utilizes" --> Coordinate_Scaler
Classifier_Application -- "Utilizes" --> Coordinate_Format_Converter
The Inference & Post-processing subsystem is central to the PyTorch_YOLOv4 project, orchestrating the object detection process from raw model outputs to refined, actionable detections. It embodies the "Inference Module" and "Utilities/Helper Functions" patterns, focusing on the final stages of the deep learning pipeline. The subsystem is primarily defined by the detect.py module, which serves as the main inference orchestrator, and key utility functions within utils/general.py that handle post-processing logic and coordinate transformations.
The primary entry point for executing the object detection inference pipeline. It manages the overall flow, from loading necessary configurations (like class names) to initiating model predictions and coordinating post-processing steps.
Related Classes/Methods:
Responsible for loading and preparing the human-readable class names that correspond to the model's output categories. This is crucial for interpreting the model's numerical predictions into meaningful object labels.
Related Classes/Methods:
Implements the Non-Maximum Suppression algorithm, a critical post-processing step that filters out redundant or overlapping bounding box detections, ensuring only the most confident and distinct detections are retained for each object.
Related Classes/Methods:
Calculates the Intersection over Union (IoU) metric between pairs of bounding boxes. This metric is fundamental for determining the overlap between boxes, which is a key operation within the NMS algorithm.
Related Classes/Methods:
Converts bounding box coordinates from the (center_x, center_y, width, height) format to the (top_left_x, top_left_y, bottom_right_x, bottom_right_y) format. This conversion is often required for various post-processing steps and visualization.
Related Classes/Methods:
Adjusts bounding box coordinates from the model's internal input resolution back to the original dimensions of the input image. This ensures that the detected bounding boxes are correctly positioned and sized on the source image.
Related Classes/Methods:
Applies an optional secondary classifier to further refine the labels or confidence scores of detected objects. This can enhance the accuracy and specificity of the detection results.
Related Classes/Methods: