How feiyull/tensorrt-alpha works
TensorRT-Alpha is a collection of CUDA/TensorRT computer-vision inference examples, not a single model service. Its principal workflow accepts batched images or video frames, preprocesses them, runs a TensorRT engine, decodes and suppresses detections, then displays or saves results. The repository also includes segmentation, pose, face detection, and saliency examples. README statements describe the shared processing stages; sampled code confirms several model-specific implementations and the YOLO base class. Some model connections are dashed where their application call paths were not excerpted.
Main entry points: yolo.cpp YOLO runtime base · app_yolov5.cpp YOLO applications · yolov8_pose.cpp YOLOv8 pose · yolov8_seg.cpp YOLOv8 segmentation
Components 17
Inference Runtime
- YOLO runtime base
utils/yolo.cpp - TensorRT engine
- Display or save
utils/utils.cpp
Object Detection
- YOLO applications
yolov5/app_yolov5.cpp - YOLO model variants
yolov3 - YOLOv8 detector
yolov8/yolov8.cpp - YOLO-NAS detector
yolonas/yolo_nas.cpp - EfficientDet
efficientdet/efficientdet.cpp
Specialized Vision
- YOLOv8 pose
yolov8-pose/yolov8_pose.cpp - YOLOv8 segmentation
yolov8-seg/yolov8_seg.cpp - Human segmentation
pphumanseg/pphunmanseg.cpp - Face detection
libfacedetection/libfacedetection.h - Saliency detection
u2net/u2net.h - Task decoders
pphumanseg/decode_pphunmanseg.h
CUDA Utilities
- CUDA preprocessing
utils/kernel_function.h - YOLO decode and NMS
utils/yolo.cpp
Other components
- Image or video source
Connections 25
- Image or video source → YOLO applications: provides frames
yolov5/app_yolov5.cpp - YOLO applications → YOLO runtime base: uses runtime
- YOLO applications → YOLO model variants: selects model
- YOLO applications → YOLOv8 detector: selects model
- YOLO applications → YOLO-NAS detector: selects model
- YOLO runtime base → CUDA preprocessing: preprocesses batches
- YOLO runtime base → TensorRT engine: executes engine
- YOLO runtime base → YOLO decode and NMS: decodes detections
utils/yolo.cpp - YOLO runtime base → Display or save: returns results
- YOLO model variants → YOLO runtime base: inherits runtime
- YOLOv8 detector → YOLO runtime base: inherits runtime
- YOLO-NAS detector → YOLO runtime base: inherits runtime
- EfficientDet → CUDA preprocessing: preprocesses batches
efficientdet/efficientdet.cpp - Image or video source → EfficientDet: provides frames
- Image or video source → YOLOv8 pose: provides frames
- Image or video source → YOLOv8 segmentation: provides frames
- Image or video source → Human segmentation: provides frames
- Image or video source → Face detection: provides frames
- Image or video source → Saliency detection: provides frames
- YOLOv8 pose → YOLO runtime base: inherits runtime
yolov8-pose/yolov8_pose.cpp - YOLOv8 segmentation → Task decoders: uses decoder
yolov8-seg/yolov8_seg.cpp - Human segmentation → Task decoders: uses decoder
pphumanseg/pphunmanseg.cpp - YOLOv8 pose → YOLO decode and NMS: applies NMS
yolov8-pose/yolov8_pose.cpp - YOLOv8 segmentation → YOLO decode and NMS: applies NMS
yolov8-seg/yolov8_seg.cpp - Image or video source → YOLO model variants: provides frames