Decentralized AI Data Labeling
Last updated
Last updated
As a full-stack AI data solutions provider, the core of the TAGGER system is designed to integrate data rights confirmation, AI-driven encryption and decryption, smart contract technology, and an AI-assisted dataset annotation system that rivals the precision of industry experts.
The data annotation module of TAGGER, as illustrated in Figure 5, operates as follows:
Application for Annotation Tasks: Users holding dataset ownership NFTs apply for annotation tasks on the platform.
Task Generation and Matching: The platform then randomly generates annotation tasks and matches the data index with the user's request, ensuring efficient task allocation.
Data Decryption for Workers: Once Web 3 data workers accept and engage in the annotation tasks, the platform employs a dynamic mixed chaotic decryption algorithm to decrypt the data assigned to them. This ensures secure access to the relevant data for the task at hand.
Graphical Annotation Tools and AI Assistance: Workers use graphical annotation tools to annotate the dataset. During the annotation process, the platform leverages a target detection algorithm to automatically identify and mark specific regions of interest within the data, reducing the need for extensive manual annotation—especially in image datasets. The expert AI annotation assistant further aids workers by improving the accuracy of annotations and continuously monitoring the results in real-time to prevent ineffective or incomplete annotations.
At the core of TAGGER's annotation module lies the AI-assisted annotation technology, which integrates general auxiliary tools with an expert knowledge base across various industries. This system enables workers to accurately annotate professional datasets without requiring specialized knowledge.
The platform employs optimized image encoders, prompt encoders, and mask decoders, powered by the knowledge base of industry experts, to minimize the annotation range for datasets. It simultaneously presents workers with the precise target content that requires annotation, effectively preventing inaccurate or ineffective annotations.
Additionally, the platform features an advanced pixel recognition module, which supervises and evaluates whether the annotated range is sufficient. This module enhances the automation of the platform while safeguarding the interests of data annotation task publishers, ensuring that annotations meet the required standards for high-quality AI dataset creation.
Figure 7 illustrates the core system’s architecture and the workflow of the AI-assisted annotation system.
The structure of TAGGER's automated dataset annotation engine is composed of several key components designed to optimize the accuracy and efficiency of dataset annotation:
Image Encoder: The platform utilizes a Vision Transformer (ViT) pre-trained with Mean Absolute Error (MAE) to handle high-resolution inputs adaptively, minimizing distortion while preserving image integrity.
Prompt Encoder: The platform supports two types of prompts—sparse and dense—based on the specific dataset requirements. Position encodings are employed to represent points and boxes, which are incorporated into the learning embeddings for each prompt type. In addition, a pre-trained text encoder from CLIP is used to represent free text. Dense prompts, such as masks, leverage convolutional embeddings added to image embeddings.
Mask Decoder: The mask decoder maps image embeddings, prompt embeddings, and output tokens into masks. It is composed of a Transformer decoder module followed by a dynamic mask prediction head to generate precise mask outputs.
Ambiguity Resolution: To resolve ambiguities in the output, this module averages multiple valid masks when ambiguous prompts are provided, ensuring the most accurate result is delivered.
Loss and Training: The platform supervises mask prediction using a linear combination of focal loss and dice loss, enhancing the performance of the model. It further incorporates geometric prompts in the training of promptable segmentation tasks to improve prediction accuracy.
The expert knowledge base that supports the automated dataset annotation engine is structured according to the platform’s planned data types and industry-specific classifications. This knowledge base builds diagnostic analysis models tailored to specific scenarios, employing deep learning algorithms to analyze the characteristics of the original datasets. By detecting target edges and features in these datasets, the system enhances the recognition accuracy of the automated annotation engine, ensuring precise and high-quality annotations.
