Visual Analysis

Visual Analysis

Minivision Tech provides you with excellent performance and safe and reliable technical services.

Robust Few-Shot Learning Technology

When implementing algorithms, the application scenarios are complex, diverse and highly random. To respond quickly to algorithm requirements, the company relies on ‘Robust Few-Shot Learning Technology’ to achieve efficient and rapid model customization, solving the problem of long-tailed distribution of data in target scenarios.

图片
  • 去水纹人脸比对识别

    In the field of de-watermarking face matching recognition, the company uses few-shot learning technologies to quickly realize the landing. The algorithms are widely used in payment and financial authentication.

    1/3
  • 2020年初即研发戴口罩人脸识别算法

    The first face recognition algorithm for people wearing masks was invented in early 2020, with 97% accuracy. This has been widely reported in national and international media, including Heart of Machine and The Times.

    2/3
  • 联合南京理工大学杨健教授团队完成“图像复原与稳健识别的理论与方法”

    Together with Prof. Jian Yang's team at Nanjing University of Science and Technology, we have completed "Theory and Methods of Image Restoration and Robust Recognition", which won the First Prize of Science and Technology Progress of Jiangsu Province in 2021. We systematically solve the singularity problem in the one-shot training process to improve the robustness of the algorithm in application scenarios.

    3/3

Computer Vision Algorithms

  • Recognition, Clustering, and High-Speed Tracking Technology Based on Multimodal Representations.
  • Fine-grained structured analysis technology for large scale video
  • 3D Computer Vision Technology
  • AI generation, authoring and AR interaction technology
  • Algorithms and engines for large-scale, high-precision face recognition in complex scenarios

    Algorithms and engines for large-scale, high-precision face recognition in complex scenarios

    Face recognition is affected by various factors such as lighting, angle, age and ethnicity, resulting in matching deviations due to scene heterogeneity. To address this, a face recognition algorithm based on ‘Scale-Transferrable’ and attention mechanisms has been developed. Predictions are made based on multi-layer features and are fused. Based on the Scale-transfer Layer proposed by the company, recognition accuracy in complex scenes is effectively improved with almost no increase in computational complexity. Under complex practical test conditions, the accuracy reached 99.8%.

    1. »In 2020, the company achieved an overall ranking of 7th in the NIST International Face Recognition Vendor Test (FRVT) and 2nd in the highly demanding Wild Test (open scene).

    2. »In 2020, the company's intelligent face recognition terminal (MV-PT1A-R2) received certification reports from the Third Research Institute of the Ministry of Public Security.

    3. »In 2021, the company's intelligent face recognition terminal (MV-PT2A-X-X) received certification reports from the Third Research Institute of the Ministry of Public Security.

  • Fast Face Retrieval

    Fast Face Retrieval

    By exploiting the fast face retrieval capabilities of ultra-large face databases and the optimized feature indexing technology, the company has reduced face retrieval time to less than one second for face databases up to 100 million in size.

  • Live authentication technology

    Live authentication technology

    This technology supports both cooperative and non-cooperative detection methods. The living algorithm can provide feedback results in milliseconds and has high accuracy. In complex scenarios, the true user acceptance rate reaches 99% and the anti-spoofing capability is over 99.9%, effectively countering various types of attacks such as photos, masks, video replays and mannequins.

    1. »In 2020, the open-source, RGB-based, industrial-grade algorithm for silent liveness detection received widespread industry attention and has now been awarded 821 stars.

    2. »In 2020, Minivison Tech achieved enhanced liveness authentication for UnionPay through BCTC. Based on RGB and near-infrared multimodal data, the company developed a high-security level face recognition algorithm for payment authentication.

  • Fine-grained structured analysis technology for large scale video

    Fine-grained structured analysis technology for large scale video

    Based on Minivison's core algorithm development platform and rich experience in video structuring technology, we already had the capability of rapid R&D, production and practical application of industrial-grade multi-scene video structuring algorithms. These algorithms have been extensively applied in various scenarios such as smart cities, communities and schools. They include more than 200 types of algorithms, including action detection such as crowd gathering and fighting, construction vehicle detection, roadside stall detection, floating object detection, objects thrown from tall buildings, and tethered dog detection.

    1. »The algorithm for detecting objects thrown from tall buildings in community scenarios was selected by Huawei Mall's SDC.

    2. »More than 30 video structuring algorithms for community scenarios have been applied in the smart community transformation of more than 200 neighborhoods in Jiangning District, Nanjing.

    3. »We are the video structuring algorithm provider for China Mobile's ‘Qianliyan’ project, providing over 20 algorithms and analyzing real-time video streams from over 1000 cameras.

    4. »Our algorithms are compatible with more than 20 home-grown AI chips, enabling algorithms to run on various edge devices.

    5. »In 2022, our algorithm took first place in the MOT Challenge in 8 evaluation metrics, with MOTA and HOTA as the core metrics.

  • Depth estimation technology

    Depth estimation technology

    The technology does not require expensive depth information sensors, and can use low-cost ordinary cameras to perform complex ranging tasks and estimate the depth of the scene.

  • Human pose estimation technology

    Human pose estimation technology

    Using a standard monocular camera, three-dimensional position information of human skeletal key points is extracted from the video frame in real time, enabling three-dimensional reconstruction of the human skeleton. It also allows precise analysis of joint movement angles, velocities and other parameters, achieving the effect of motion capture.

  • High-precision 3D facial reconstruction technology

    High-precision 3D facial reconstruction technology

    The technology is capable of real-time detection of 53,212 3D key points on the human face with a normalized mean error of less than 2.8%. Based on the 3D fine modelling of faces using key points, it can accurately capture various subtle facial expressions.


    »In the early stages of the 2020 epidemic, we used AI generation technology to simulate data due to a severe lack of masked face data. The generation technology produced training data with consistent mapping of complex scenes, which helped our masked face recognition algorithm to go live in a short period of time.

  • Techniques for image generation and data simulation

    Techniques for image generation and data simulation

    We can augment a limited and expensive real-world dataset and use its samples as training samples to improve the performance of production models and reduce the cost of model development. This also has a wide range of applications in interactive entertainment and artistic creation.

  • Image stylization technology

    Image stylization technology

    It can convert images into specific style types or imitate the artistic style of masters for artistic creation, thereby achieving high-quality photo and video enhancement that can enrich content details and add special effects, etc.


    »After being partially open sourced on Github, our portrait cartoonization technology has attracted a lot of attention from the industry. It has long been at the top of the hot list in its field and has accumulated more than 3,200 stars.