LinkedAI is a Y-Combinator funded AI startup platform building highly accurate training datasets for computer vision-related use cases using machine learning at scale to reduce data annotation and labelling time. The company was launched in 2018 and is headquartered in Sunnyvale, California. Founded by Paula Villamarin(CEO) and Diego Parra(CTO). LinkedAI specialises in Computer vision, Data labelling, Data tagging, NLP, generating synthetic data. It allows team collaboration and labelling quality checks by professional annotators. A workflow of LinkedAI is shown below:
Auto labelling with ML-powered services reduces both cost and time. The platform has both in house trained labellers and crowdsourcing for larger projects.
Features – LinkedAI has a wide range of options in which for annotation tools like bounding boxes, polygons, lines, semantic segmentation and landmarks for various use cases.
Bounding Boxes are one the very basic type of labelling methods. It is drawn around the object of interest. This is mostly used in object detection models for self-driving cars and other use cases.
Polygons are a more precise form of data labelling technique which use pixel-precise to determine the object boundaries. This creates a clear shape and blob around the required object. Polygons have proven to be better than boxes as they avoid any unnecessary thing present with the object of interest. This doesn’t confuse the model and gives better predictions. It is used mostly in object localization.
Semantic segmentation or Pixel-level annotation can group together the pixels that have similar attributes. These differentiate each object in an image or frame. It is used for models in detection and localization of specific objects. Unlike polygon annotation that focuses on only objects of interest, full semantic segmentation gives a complete overview of every pixel of the scene in the frame.
Landmark Annotation performs point-to-point variations in small objects, it helps to mark each point motion in the targeted object. Key points can help in understanding facial recognition or gestures. These are also used in estimating poses with accuracy.
Lines Annotation is used to train self-driving vehicle models for lane detection.
Image categorization classify and label images according to visual content
LinkedAI has an amazing feature of generating synthetic data to enhance training datasets. This is called Flip that uses transformers to generate new 2D images from small batches of objects and superimposed on some background images.
LinkedAI Platform – users can try for free with their own images or sample images to check the annotation tools made available by LinkedAI.
In the side panel images can be uploaded under dataset or project section, categories and labels can be managed from here itself. The toolbar contains filters (Restore, Opacity, Saturation, Brightness, Contrast, and Grayscale), change modes and auto labels.
The annotated data can be downloaded in JSON or CSV format.
Example of a JSON:
[ { "name": "2D Box", "color": "blue", "type": "bounding_box", "pos": { "x": 419.0453163419915, "y": 164.2535461607743, "w": 50.61413969335605, "h": 25.597955706984667 } },
{ "name": "Polygon", "color": "green", "pos": [ { "x": 663.1877801517137, "y": 200.39595155645543 }, { "x": 740.7573429001523, "y": 198.84456030148664 }, { "x": 759.3740379597774, "y": 236.07795042073707 } ], "type": "polygon" } ]
Example of a CSV:
name,color,type,x,y,w,h "Bounding Box","blue","bounding_box","550.823","322.332","134.195","98.513344"
The next section is Canvas where labels are added to images. There are certain keyboard shortcuts also available for quick access to options.
Companies
CocaCola, Kiwibot, River, LiciMatic, tus datos, servinformacion, Transport Systems, Asimetrix, Universidad de Los Ángeles.