Newsletter

Ever since we launched VDP in August, we've been actively collecting feedback from our community. Along the way we've realised that as a general ETL infrastructure, VDP is in fact capable of processing all kinds of unstructured data.

We should not limit its usage to only visual data but to more versatile data. In addition, the term, Data Preparation, has been misleading for users often thinking it has to do with only data labelling, cleaning, or wrangling. In our vision, while VDP should involve data preparation in its MLOps practice, it should not be conceptually confined to only data preparation. VDP does more than that and is focused on the overall effectiveness of the unstructured data ETL with a data-centric paradigm. In spite of the high complexity of the underneath data processing, the end form of a VDP infrastructure is nothing but a data pipeline. The term Data Pipeline is more precise to capture the core concept of VDP, hence renamed as "Versatile Data Pipeline".

VDP now supports the instance segmentation task! It is a Vision task to detect and delineate multiple objects of pre-defined categories in an input image.

To build an instance segmentation data pipeline, check out the Streamlit example and import the Mask R-CNN model from the GitHub repo instill-ai/model-instance-segmentation-dvc.

👉 Learn more

Besides image classification, object detection and keypoint detection, VDP now supports the Optical Character Recognition (OCR) task!

It is a computer vision task to detect bounding boxes containing text and to process typed or handwritten text within each bounding box into machine readable text. To build an OCR data pipeline, try to import the model from the GitHub repository instill-ai/model-ocr-dvc.

👉 Learn more

A few months ago, we announced that we're building VDP, an open-source data tool to streamline the end-to-end visual data processing pipeline:

• Extract unstructured visual data from pre-built data sources such as cloud/on-prem storage, or IoT devices
• Transform it into analysable structured data by Vision AI models
• Load the transformed data into warehouses, applications, or other destinations

VDP integrates with Airbyte to provide open-source data connectors.

By leveraging the pre-built and ready-to-use data connectors, VDP is the single point of visual data integration. You can sync visual data from anywhere into a centralised warehouse and focus on gaining insights across all your data sources, instead of maintaining pipelines.

👉 Learn more

VDP integrates with the best ML tools to make importing models super easy.

No need to change how you manage your models. No matter the model is on your computer, in a GitHub or Hugging Face repository, or in cloud storage managed by version control tools like ArtiVC or DVC, import & deploy it with one click 🖱️.

👉 Learn more

VDP supports data pipelines for diverse scenarios - no need to build from scratch

👉 Learn more

VDP solves popular vision tasks out of the box:

• Image classification
• Object detection
• Keypoint detection
• OCR (coming soon!)
• The list is growing…🌱

For each task, you can import corresponding STOA models into VDP to use in the data pipelines. The prediction results are automatically converted to a standardised format ready for future analysis.

👉 Learn more

Our CEO Ping-Lin joined a panel in this year's Galaxy Summit with Taboola's Vice President Ning Ning Wu, Gojek's Vice President Pulkit Khanna, Canner Data's CEO Howard Chi and Eslite's Vice President Frank Zhang to talk about the challenges of unstructured data processing and Vision AI in the modern data stack.

Instill AI is dedicated to making unstructured visual data analysable because only analysable data can create value. We build the open-source VDP to streamline the end-to-end visual data processing across different stakeholders.

📺 Watch the panel stream here.

• Ingest unstructured visual data from data sources such as data lakes or IoT devices;
• Transform visual data to meaningful structured data representations by Vision AI models;
• Load the structured data into warehouses, applications, or other destinations.