HumanPose_Face_Analysis

This project contain face,huamn-pose detect(yolo,rtmdet,etc) and recognition(arcface,rtmpose,etc). Try to provide inference on onnx,tensorRT,rknn...

Update Log

2025/06/26 FaceRecognition(insightface) can use bmodel(sophon算能) to inference.
2025/06/21 Now, we support tensorRT(10.x) for RTX 50 Series(BLACKWELL).
2025/04/03 You can download onnx models by google dirver(https://drive.google.com/drive/folders/1DfTw0aEpuEyXpo7XJXIvCzDTuZ-wNOy8?usp=drive_link). See details in ai/models/readme.
2025/01/23 Add some new feature(Ocr,TTS,YOLO_obb). (Warning, this tts model is only support chinese! | TTS目前仅仅支持中文，且encoder还不能转化为engine，静候更新x。X)
2024/10/31 Support Onnx and tensorRT(8.6.1) inference.

Requirements

python>=3.8
torch==2.1.1
onnxruntime-gpu==1.18.0
tensorrt==8.6.1

If you are using RTX 50 Series, you are better to install tensorrt 10.8+, python 3.10+, cuda12.x and torch 2.7+.

In this project, u dont need to install insightface or ultralyitcs or paddlepaddle or mmpose!

USAGE

There will briefly introduce how to use it.

Face

Face, check face_reco.py. See the detail of the class FaceRecognition. And 'main' is the demo. Also provide single choose for face det,recognition,direction,quality(see the file for how to use).

Model ### use face_det_10g for face detection;use face_w600k_r50 for face feature extraction;use face_quality_assessment for face quality assessment.
Usage ###

# init model
det_path="your_path_face_det_10g"
reg_path="your_path_face_w600k_r50"
quality_path="your_path_face_quality_assessment"
face_recognitio = FaceRecognition(det_path=det_path,
                                  reg_path=reg_path,
                                  quality_path = quality_path,
                                  gpu_id=0)
# init data
group_embedding = np.load("your_path_group_embedding")
group_id = np.load("your_path_group_id")
# detect face
frame = cv2.imread("img_path")
results = face_recognitio.detect(frame, quality=False)  
'''
results is a list of dict, each dict is a face, keys are follows:
bbox is a int32 numpy array of [x1,y1,x2,y2]
kps is a int32 numpy array of [x1,y1,x2,y2,x3,y3,x4,y4,x5,y5], representing the five facial landmarks
embedding is a float32 numpy array of 1x512
when quality is True, error_code and error_message will be returned, detail see the file
'''
# match face
for result in results:
    # frame = cv2.rectangle(frame, (result['bbox'][0], result['bbox'][1]), (result['bbox'][2], result['bbox'][3]),(0, 0, 255), 2) ### if u wanna draw the face box, u can use this code 
    in_embedding.append(result['embedding'])
    match_info = face_recognitio.match_feature(result['embedding'], group_embedding, group_id, thre=0.7)
'''
match_info is a list of dict, each dict is a person, keys are follows:
person_id is a string, refer to group_id.
similarity is a float, show the max similarity.
'''

Pose

Pose, check mmpose.py. See the detail of the class TopDownEstimation. And 'main' is the demo. Also provide single choose for det(yolo,rtmdet),pose(rtmpose),track(deepsort)

Segment

Beside, there are some segment methods for human, such as PSPNet,yolo-seg,detins. But it still need valid, so i didnt package it.

Ocr

Ocr, check ppocr.py. See the detail of the class PaddlePaddleOcr. And 'main' is the demo.

TTS

Tts, check t2s.py. See the detail of the class Text2Speech. And 'main' is the demo. (Only cn)

YOLO_obb

YOLO_obb, use yolo_obb.py. See the detail of the class YOLO_OBB. And 'main' is the demo.

License

This repository is licensed under the Apache-2.0 License.

Reference

Insightface https://github.com/deepinsight/insightface
FaceImageQuality https://github.com/pterhoer/FaceImageQuality
MMpose https://github.com/open-mmlab/mmpose
rtmlib https://github.com/Tau-J/rtmlib
deep_sort_pytorch https://github.com/ZQPei/deep_sort_pytorch
PaddleOCR https://github.com/PaddlePaddle/PaddleOCR