Supported Modules#

The tables below list the detectors, trackers, and ReIDers integrated in pyppbox for the PoseTReID framework (read more here). Each integration exposes the features needed by the pipeline, rather than every feature of the original project. For example, YOLO_Classic uses official .weights files for detection; it does not train YOLO models.

Config names are the values used in dt_name, tk_name, or ri_name. CPU/GPU entries describe backend capabilities. Device controls and automatic selection differ by module; see Devices for editable dropdowns, presets, runtime requirements, and the OpenCV/TensorFlow automatic-device rows.

Supported Detectors#
General Name	Config Name	Details
YOLO Classic	`YOLO_Classic`	Integration: built-in OpenCV DNN detector Models: `.weights` files for V2, V3 and V4 Runtime: CPU or GPU; OpenCV selects the device automatically
YOLO Ultralytics	`YOLO_Ultralytics`	Integration: linked ultralytics package Models: `.pt` files for V3, V5, V8, V9, v10, v11, v12, v26 Runtime: PyTorch; configurable device
RF-DETR	`RF_DETR`	Integration: optional `requirements/rfdetr.txt` on Python 3.10+ Models: RF-DETR Nano, Small, Medium, and Large Runtime: PyTorch; configurable device Example: Example 15: Use RF-DETR detection and SORT tracking
Ground truth	`GT`	Integration: built-in ground-truth reader Input: detections and identities from a ground-truth text file Runtime: CPU Example: Example 10: Use GT (Ground-truth) as a detector

For GT, set both tracker and reider to "None" to preserve the supplied identities automatically. With other stages enabled, pyppbox.ppb.mt.MT.forceFullGTMode() overrides detection-only GT behavior until the next module selection.

Supported Trackers#
General Name	Config Name	Details
Centroid	`Centroid`	Integration: built-in centroid tracker Runtime: CPU
SORT	`SORT`	Integration: embedded SORT repo Runtime: CPU
DeepSORT	`DeepSORT`	Integration: embedded DeepSORT repo Runtime: TensorFlow appearance encoder; CPU assignment Device: TensorFlow selects devices automatically

Supported ReIDers#
General Name	Config Name	Details
FaceNet	`FaceNet`	Integration: embedded FaceNet repo Models: MTCNN face detection, FaceNet embeddings, and an identity SVM Runtime: CPU or GPU; TensorFlow selects devices automatically Classifier training: Example 12: Retrain GTA V classifiers for all supported reiders
Torchreid	`Torchreid`	Integration: linked pyppbox-torchreid package Models: OSNet-AIN, OSNet, and MLFN embeddings with an identity SVM Runtime: PyTorch; configurable device Classifier training: Example 12: Retrain GTA V classifiers for all supported reiders
ArcFace	`ArcFace`	Integration: optional `requirements/faces.txt` or `requirements/faces_gpu.txt` on Python 3.10+ Models: SCRFD alignment and local ArcFace R50 ONNX embeddings Runtime: ONNX Runtime on CPU or NVIDIA CUDA; configurable device Identity SVM: separate ArcFace classifier trained with frozen embeddings Classifier training: Example 12: Retrain GTA V classifiers for all supported reiders Examples: Example 17: Compare FaceNet, ArcFace, and AdaFace, Example 18: Recognize faces and train an ArcFace or AdaFace classifier
AdaFace	`AdaFace`	Integration: optional `requirements/faces.txt` or `requirements/faces_gpu.txt` on Python 3.10+ Models: SCRFD alignment and local AdaFace IR18 ONNX embeddings Runtime: ONNX Runtime on CPU or NVIDIA CUDA; configurable device Identity SVM: separate AdaFace classifier trained with frozen embeddings Classifier training: Example 12: Retrain GTA V classifiers for all supported reiders Examples: Example 16: Export an AdaFace embedding model to ONNX, Example 17: Compare FaceNet, ArcFace, and AdaFace, Example 18: Recognize faces and train an ArcFace or AdaFace classifier

ArcFace and AdaFace use their respective v1.6.0 data wheels for the supplied models and classifiers. See Supported ReIDers for installation, custom models, classifier compatibility, and training requirements.