Yifan Liu

Key Skills

Software

AWS SageMaker

Top Subject Matter

Self-driving

VLM in end-ti-end model

3D object detection

Top Data Types

3D Sensor

Image

Video

Top Task Types

Action Recognition

Classification

Fine Tuning

Prompt Response Writing SFT

Segmentation

Freelancer Overview

An experienced AI scientist with a strong foundation in data labeling and AI training data. Demonstrated expertise in leading the development of active learning algorithms for 3D object detection, notably utilizing the CRB framework to reduce annotation costs by 85%. Skilled in working on large-scale data preprocessing, model training, and optimization for Video NLP projects, significantly enhancing model performance in semantic understanding and video analysis. Adept in multi-label classification systems and advanced image restoration techniques, with a focus on improving model robustness and generalization through innovative degradation pipelines. Proven ability to deliver high-performance AI solutions through meticulous dataset construction, model tuning, and algorithm customization.

Labeling Experience

End-to-end large-scale Video Natural Language Processing (Video NLP)

Aws SagemakerVideoBounding Box

Participated in an end-to-end large-scale Video Natural Language Processing (Video NLP) project, responsible for data preprocessing, model training, and optimization, successfully improving the model’s semantic understanding and video analysis capabilities.

Participated in an end-to-end large-scale Video Natural Language Processing (Video NLP) project, responsible for data preprocessing, model training, and optimization, successfully improving the model’s semantic understanding and video analysis capabilities.

2024

Fashion Attributes Classification Challenges

Aws SagemakerImageBounding Box

1. Designed and implemented an efficient multi-label classification system for fashion images, aimed at precisely identifying clothing attributes in fashion photographs. The system converts the complex fashion attribute classification task into a multi-label problem, providing a detailed prediction framework capable of identifying 26 finely categorized attributes across six core categories, thereby significantly enhancing classification granularity and practical utility. 2. Conducted an in-depth comparison of classic architectures including ResNet50, DenseNet121, and VGG, ultimately selecting EfficientNet_B4 as the base model. Performed comprehensive model tuning and performance testing to ensure optimal accuracy and efficiency of the system.

1. Designed and implemented an efficient multi-label classification system for fashion images, aimed at precisely identifying clothing attributes in fashion photographs. The system converts the complex fashion attribute classification task into a multi-label problem, providing a detailed prediction framework capable of identifying 26 finely categorized attributes across six core categories, thereby significantly enhancing classification granularity and practical utility. 2. Conducted an in-depth comparison of classic architectures including ResNet50, DenseNet121, and VGG, ultimately selecting EfficientNet_B4 as the base model. Performed comprehensive model tuning and performance testing to ensure optimal accuracy and efficiency of the system.

2024 - 2024

Active learning algorithms for a 3D object detection model

Aws Sagemaker3D SensorSegmentation

Led the development of active learning algorithms for a 3D object detection model, leveraging the CRB (Conciseness, Representativeness, Balance) framework — an innovative approach tailored for point cloud acquisition. This method dramatically reduced manual annotation costs by focusing on the most informative data. Successfully achieved performance comparable to using the full dataset with only 15% of the data, thereby enhancing data labeling efficiency and overall model effectiveness

Led the development of active learning algorithms for a 3D object detection model, leveraging the CRB (Conciseness, Representativeness, Balance) framework — an innovative approach tailored for point cloud acquisition. This method dramatically reduced manual annotation costs by focusing on the most informative data. Successfully achieved performance comparable to using the full dataset with only 15% of the data, thereby enhancing data labeling efficiency and overall model effectiveness

2024 - 2024

Education

Work History