BLOG Machine Learning

Bounding Boxes for Object Detection and Classification

Mothi Venkatesh
July 20, 2018

2D Bounding Boxes is perhaps the most ubiquitous annotation type one might encouter in computer vision. As the name suggests, the annotator is asked to draw a box over the objects of interest based on the requirements of the client. Apart from object classification and localization models that could be trained using bounding boxes, they could also be used for tasks such as semantic-segmentation as weak supervision approaches. This flexible nature of bounding boxes makes its presence advantageous for training models on abstract, noisy and intuitive labels for applications such as self driving cars, furniture, clothing and satellite imagery.

Bounding Boxes have the lowest time per annotation and is also the cheapest in terms of cost.

Most common use cases we've handled:

#1 Object localization for Self-driving cars

Bounding boxes for Autonomous driving

Extensively used to train autonomous driving perception models for pedestrians, traffic signs, lane obstacles, etc. For ex: Bounding boxes can be used to annotate various fashion accessories and this is used to train visual search machine learning models.#2 Object Detection for e-commerce

Bounding boxes for fashion eCommerce

Used to train visual search machine learning models for recognition of various fashion accessories and furniture.#3 Damage detection for Insurance

Bounding boxes for vehicle insurance

Identification of car damage, roof damage or safety parameters from live world images to train machine learning models that detect the degree of damage for insurance claims.#4 Drone and Robot training

Bounding boxes for Autonomous driving

Labelled images for training smart surveillance drones and robots to identify a variety of objects.#5 Object detection for Furniture

Bounding boxes for furniture

Used to train ML for object detection in a picture of a room, bounding boxes can be drawn over the furniture such as chairs, tables, cupboards etc.