Implementation of End-to-End YOLO Models for TensorRT
This repository supports segmentation and detection models from the YOLO series - YOLO11, YOLOv10, YOLOv9, YOLOv8, and YOLOv7.
Use the repository https://github.com/levipereira/ultralytics specifically for exporting the YOLO11, YOLOv10, and YOLOv8 models.
All YOLO series models are implemented with End2End Deep Learning, incorporating three key features:
- Dynamic Shapes - TensorRT enables the creation of network resolutions different from the original exported ONNX.
- Dynamic Batch Size - Dynamically adjusts the batch size to maximize model performance according to the GPU's capabilities.
- NMS-Free - Achieved in two different ways:
- Native NMS-Free - Models natively support NMS-Free, available for some YOLOv9 models and all YOLOv10 detection models.
- TensorRT Plugins - Utilizes TensorRT EfficientNMS plugin for detection models, and EfficientNMSX and ROIAlign for segmentation models.
All detection models across the YOLO series adhere to standardized output layers:
num_det: Represents the number of detections.det_boxes: Provides the bounding boxes coordinates for each detected object.det_scores: Indicates the confidence score associated with each detected object.det_classes: Specifies the class label or category assigned to each detected object.
Similarly, instance segmentation models from the YOLO series also maintain standardized output layers:
num_det: Represents the count of detected instances.det_boxes: Provides bounding box coordinates for each detected instance.det_scores: Indicates the confidence score associated with each detected instance.det_classes: Specifies the class label assigned to each detected instance.det_masks: Provides the segmentation masks corresponding to each detected instance.