This repository contains an example of using Rekognition custom labels for Object detection. The initial dataset was copied from Context-based information generation from construction site images using unmanned aerial vehicle (UAV)-acquired data and image captioning. The dataset is licensed as CC BY 4.0.
The dataset contains a set of 1431 annotated images in jpg format, named following the regular expression [1-9][0-9]*.jpg.
Annotations consist of a collection of regions numbered from 0, defined by bounding boxes with a textual description of what was identified, generally o
- Bounding box: each region contains a
shape_attributesfield, containing the following subfields:name: usually just "rect", the shape of the box;x: region start x reference, as pixes from left (to be confirmed).y: region start y reference, as pixes from top (to be confirmed).width: width of the region in pixels.height: height of the region in pixels.
- Textual description: each region contains a
region_attributesfield, itself containing aphrasesubfield f the format<qualifier>* <object> <locative verb/preposition> <location>:qualifiercan be a color, a collection, a quantity, a temporal qualifier or others. Many qualifiers can exist.objectis the class of interest, what is to be identified.locative verb/prepositiondetermines the spacial relationship betwen theobjectand thelocation.locationis a reference area, such as "the ground", "concrete", "structure". It can have its own qualifiers.