50+ new layers
Nearmap has enhanced its machine learning capabilities significantly, resulting in the addition of over 50 new AI layers in Gen 6. This integration results in a total of 130+ AI layers, enriching our location intelligence services across all verticals to:
- Allow remote inspection and assessment
- Classify and detect damage
- Optimize workflows
- Enable fast and accurate decision-making
- Manage risks more effectively
Improved detection and identification
We have significantly improved the AI detection and identification of existing layers. We have also made substantial investments in our model that include a brand new deep learning model architecture and post processing stack, and a significantly larger training data set. With these changes, Gen 6 outperforms both Gen 5 and existing Betterview products by a significant margin. Over 75% of the Gen 5 layers and 95% of the incoming Betterview detections (AI Layer equivalents) have been improved.
New property based metrics on a fair test set see f1 score improvements anywhere from a few percentage points to increases of more than 50%. For example, the swimming pool parcel presence f1 score goes from 91% and 94% for Betterview and Gen 5 results respectively, to over 98% with Gen 6. These improvements will flow through to have a positive impact on all derived scores, such as damage classification and roof condition scores.
Improvements over Gen 5
- Additional labeling effort equivalent to over 227 person-years beyond the Gen 5 model
- Training dataset 2.5 times larger than the Gen 5 with a total of 2 million images
- Over 300% increase in the number of production layers
- Innovative Deep Learning architecture, with larger backbone to allow capacity for learning the nuances of nearly twice as many output layers.
- Completely rebuilt post processing stack for efficiency, fault tolerance and improvements to building and roof outlines.
- Many layers have been migrated from alpha or beta status to production.
Format and data changes
- By default, the AI Feature API will return Gen 5 data. If you wish to access Gen 6 AI via the API, use the API parameter
systemVersionPrefix=gen6-. - Low confidence roof score, available in Gen 5 via the AI Feature API, is not available in Gen 6 AI.
Coverage
Processing of historical surveys has been conducted on Gen 6 across all surveys in:
- USA and Canada starting from July 1, 2023
- Australia and New Zealand from July 1, 2024
Version history
gen6-glowing_grove-1.0
Version | Updates | New AI Packs | Enhanced AI Packs |
|---|
1.0 | - Original building class deprecated and a new "Building" feature class added
- Original "Flat Roof" class deprecated and a new "Flat" feature class added
- Original "Skylight" class renamed to "Translucent Roofing"
- Updates to the Enhanced Building Outline algorithm
- Several Alpha and Beta layers moved to Production
| | Experimental |
Details
- The original building class (which was identical to roof) has been deprecated, a new Building class has been added. The original "Building" feature class (ID `a2e4ae39-8a61-5515-9d18-8900aa6e6072`) has been flagged as deprecated and will be removed completely in Gen 7. "DEPRECATED" will appear in the description names returned by the API.
- The "Building" feature class (ID `1878ccf6-46ec-55a7-a20b-0cf658afb755`) has been de-coupled from the "Roof" class. It is now semantically distinct and includes areas such as rooftop car parks, un-roofed decks attached to buildings and so on. The definition of each of these attributes is available in the Help Center.
- Original "Skylight" class is now called "Translucent Roofing". This has been done to better reflect the behavior of the model, and does not constitute a change to the definition. Additionally, a new class named Skylight has been added, which includes windows or glass panels on roofs and does not include glass enclosures.
- Original "Flat Roof" class has been deprecated and a new "Flat" feature class has been added. We recommend using this new layer for Flat roof as we found it to be more performant than the old Flat roof layer.
Enhanced Building Outline algorithm
Updating the Enhanced Building Outline algorithm (used on both roofs and buildings) to provided higher quality outlines of complex structures.
New AI Packs
- Building Structures: The Building Structures AI pack comprises a wide array of classes designed to assist in identifying and comprehending details pertaining to various structures.
- Roof Shape: The Roof Shape pack includes a set of layers detecting the presence of the distinctive forms and structural design elements of roof areas.
- Roof Materials: The Roof Materials AI Pack consists of materials commonly used in flat roof construction. This pack can be used to identify and analyze different flat roof materials efficiently. Whether studying durability, conducting cost assessments, or planning roof maintenance projects, the Flat Roof Materials AI Pack provides valuable insights into roofing elements.
- Commercial Roof Objects: The Commercial Roof Objects pack includes a set of objects commonly found on the large, flat or low-sloped roof areas of commercial buildings.
- Advanced Roof Condition: The Advanced Roof Condition pack provides granular, detailed information about the state of repair of a roof and and roof covering materials - both longer-term degradation, as well as short term damage due to catastrophe events.
- Roof Overhang: The Roof Overhang pack contains attributes of objects that commonly extend over roof areas.
- Pavement Markings: The Pavement Markings pack focuses on painted markings on road surfaces, and provides a level of breakdown into key categories.
- Pavement Condition: The Pavement Condition pack provides insights into different conditions of an existing pavement. It aids in research to pinpoint imperfections in asphalt surfaces, including damage, cracks, disintegration, repairs, or significant staining.
- Yard Objects: The Yard Objects AI Pack is comprised of items commonly found in outdoor spaces.
- Utilities: The Utilities pack contains objects and attributes related to essential services that provide basic amenities to households and businesses.
Enhanced AI Packs
Experimental
The Gen 5 Experimental class has been updated and layers in the pack pertaining to pavement condition have to been moved to the Pavement Condition pack.
Gen 6 AI Layers
The Gen 6 status attribute results in three categories of layers:
- Production Layers: Semantically stable layers, which are suitable for general use and for which have been tested against a use case and determined to perform well.
- Beta Layers: Many of these layers are high quality, but are currently under customer testing for specific use cases.
- Alpha Layers: These layers are typically more experimental in nature and may not have been tested against a specific use case.
To evaluate a pack with Alpha or Beta layers in it, set the AI Feature API parameter "beta=true" or "alpha=true". These layers will also be available for offline delivery and are distinguished through the status attribute. For more information, see the section AI Layer Lifecycle in the About AI Layers article.