English | 中文

OpenIVA is an end-to-end intelligent video analytics development toolkit based on different inference backends, designed to help individual users and start-ups quickly launch their own video AI services.
OpenIVA implements varied mainstream facial recognition, object detection, segmentation and landmark detection algorithms. And it provides an efficient and lightweight service deployment framework with a modular design. Users only need to replace the algorithm model used for their own tasks.

1. Common mainstream algorithms

• Provides latest fast accurate pre-trained models for facial recognition, object detection, segmentation and landmark detection tasks

2. Multi inference backends

• Supports TensorlayerX/ TensorRT/ onnxruntime

3. High performance

• Achieves high performance on CPU/GPU/Ascend platforms, achieve inference speed above 3000it/s

4. Asynchronous & multithreading

• Use multithreading and queue to achieve high device utilization for inference and pre/post-processing

5. Lightweight service

• Use Flask for lightweight intelligent application services

6. Modular design

• You can quickly start your intelligent analysis service, only need to replace the AI models

7. GUI visualization tools

• Start analysis tasks only by clicking buttons, and show visualized results in GUI windows, suitable for multiple tasks

• i5-10400 6c12t
• RTX3060
• Ubuntu18.04
• CUDA 11.1
• TensorRT-7.2.3.4
• onnxruntime with EPs:
  • CPU(Default)
  • CUDA(Manually Compiled)
  • OpenVINO(Manually Compiled)
  • TensorRT(Manually Compiled)

Run
python test_landmark.py
batchsize=8, top_k=68, 67 faces in the image

• Face detection
  Model face_detector_640_dy_sim

  onnxruntime EPs	FPS	faces per sec
  CPU	32	2075
  OpenVINO	81	5374
  CUDA	105	7074
  TensorRT(FP32)	124	7948
  TensorRT(FP16)	128	8527

• Face landmark
  Model landmarks_68_pfld_dy_sim

  onnxruntime EPs	faces per sec
  CPU	69
  OpenVINO	890
  CUDA	2061
  TensorRT(FP32)	2639
  TensorRT(FP16)	3131

Run
python test_face.py
batchsize=8

• Face embedding
  Model arc_mbv2_ccrop_sim

  onnxruntime EPs	faces per sec
  CPU	212
  OpenVINO	865
  CUDA	1790
  TensorRT(FP32)	2132
  TensorRT(FP16)	2812

Run
python test_yolo.py
batchsize=8 , 4 objects in the image

• YOLOX objects detect
  Model yolox_s(ms_coco)

  onnxruntime EPs	FPS	Objects per sec
  CPU	9.3	37.2
  OpenVINO	13	52
  CUDA	77	307
  TensorRT(FP32)	95	380
  TensorRT(FP16)	128	512

  Model yolox_m(ms_coco)

  onnxruntime EPs	FPS	Objects per sec
  CPU	4	16
  OpenVINO	5.5	22
  CUDA	46.8	187
  TensorRT(FP32)	64	259
  TensorRT(FP16)	119	478

  Model yolox_nano(ms_coco)

  onnxruntime EPs	FPS	Objects per sec
  CPU	47	188
  OpenVINO	80	320
  CUDA	210	842
  TensorRT(FP32)	244	977
  TensorRT(FP16)	269	1079

  Model yolox_tiny(ms_coco)

  onnxruntime EPs	FPS	Objects per sec
  CPU	33	133
  OpenVINO	43	175
  CUDA	209	839
  TensorRT(FP32)	248	995
  TensorRT(FP16)	327	1310

• Multi inference backends

  • onnxruntime
    • CPU
    • CUDA
    • TensorRT
    • OpenVINO
  • TensorlayerX
  • TensorRT

• Asynchronous & multithreading

  • Data generate threads
  • AI compute threads
  • Multifunctional threads
  • Collecting threads

• Lightweight service

  • prototype

• GUI visualization tools

• Common algorithms

  • Facial recognition

    • Face detection

    • Face landmark

    • Face embedding

  • Object detection

    • YOLOX

  • Semantic/Instance segmentation

  • Scene classification

    • prototype

• Data I/O

  • Video decoding
    • OpenCV decoding
      • Local video files
      • Network stream videos
  • Data management
    • Facial identity database
    • Data serialization