Human-powered pixel-level image segmentation and annotation by API.

We apply machine learning in a multitude of ways to make our workforce more efficient and ensure impeccable quality and accuracy.

Object recognition is a diverse and complex problem which underlie many different mechanisms. The main objective is to produce algorithms for different image segmentation problems that can scale nicely with data and require the least amount of feature engineering to achieve its goal. Deep learning methods fit particularly well with this motivation and have proven to be very efficient in different domains, such as natural language processing, speech recognition and computer vision.

Different algorithms to tackle three different large-scale segmentation problems:

(a) image calsification

(b) detection with boxes

(c) semantic segmentation

(d) detection with segments

Semantic Segmentation.

Semantic segmentation also known as scene labeling, is the task of labeling each pixel of an image with the category it belongs to. This is a challenging task as it consists of solving both segmentation and multi-label recognition at once. Another challenge is the large-scale nature of the task: simply labeling one thousand 320 £ 240 images with a computer algorithm already corresponds to producing over 76 million pixel labels. Paradoxically, in a database of 1000 images, most object classes occur only few times. In addition, the per-class pixel distribution is often quite unbalanced: some objects like ‘sky’ tend to cover much more pixels than other objects like ‘moon’. To add a level of difficulty, hand-labeling images are very costly (as it requires segmenting objects at pixel level).

Segment Object Proposal Generation.

Object proposal algorithms aim to find diverse regions in an image which are likely to contain objects (independent of its category). As in semantic segmentation, these algorithms output a set of masks from an image. Unlike semantic segmentation, instead of generating one label for each pixel of an image, the interest of object proposal is to generate a set of regions that are likely to fully contain objects. An ideal proposal method should possess three key characteristics: (i) high recall (i.e. proposed regions should contain the maximum number of possible objects), (ii) high recall should be achieved with a minimum number of regions as possible and (iii) the proposal regions should match the object as accurately as possible. Object proposals have many applications in computer vision, e.g., object detection, weakly supervised learning, class-agnostic detection.

Object Detection with Segments.

This problem aims at finding regions in an image that fully delineates an object as well as giving a label to each region. Object proposals play a key role in modern object detection problems. State-of-the-art object detection methods consist of two-phases: (i) a rich set of object proposals is generated and (ii) a powerful classifier (usually a CNN) is applied to each proposal. Using this pipeline, strong segment object proposal can be coupled with a classifier to deal with object detection with boxes.