Talk:Microgrants/Microgrants 2020/Proposal/Detected Features Project

Unclear POSM

Is "POSM device" is related to Portable OpenStreetMap? (note - this is not an official review, I am a random mapper) Mateusz Konieczny (talk) 15:52, 11 May 2020 (UTC)

Yes, POSM is related to Portable OpenStreetMap. The cost budgeted in the proposal is the cost required to build a portable server (Inter NUC).

"software to successfully import into OSM features"

What exactly this software is supposed to be doing? Detecting features on aerial images? Guessing geometries? Conflating with OSM data? Being a specialized editor allowing to review such data?

How it will be better than already existing tools? For example RapiD?

Are you aware about Import/Guidelines?

(note - this is not an official review, I am a random mapper)

Mateusz Konieczny (talk) 15:52, 11 May 2020 (UTC)

The end goal of the summer intern’s software assignment will be to export already detected features to OpenStreetMap following the import guidelines specified by OSM. The proposed software will not detect any features nor will it assign any attribute to the features. The proposed software will fetch geometries and attributes of features detected by a machine learning model (another project, not associated with this proposal), convert it into an OSM readable format, and import it into OSM.

We intend to use the proposed software in countries like Afghanistan, where high resolution satellite imagery is not available. On testing RapID in Jalalabad, Afghanistan, no building features were detected by the Microsoft AI (possible due to low resolution satellite imagery). Our feature detection model will use multispectral/RGB UAV imagery if high resolution satellite imagery is not available for feature detection (also another project, not associated with this proposal). Thus, generating accurate feature detection and high quality OSM data subsequently.

• "another project, not associated with this proposal" - is it already existing or just planned? Because it is OK to import only datasets of high quality, and with not existing machine learning tool it would be extremely risky to blindly fund this project Mateusz Konieczny (talk) 23:43, 16 May 2020 (UTC)

Automatic feature detection and geometry generation as well using UAVs to collect images for a disaster risk-based decision support system are part of an ongoing project we’re involved with at Virginia Tech University. We can provide additional information about this project if requested.

"We can provide additional information about this project if requested." - for start: is it completed? Where its output data is published? How project and output data is licensed? Mateusz Konieczny (talk) 15:18, 17 May 2020 (UTC)

The automatic feature detection software is being developed with a beta version expected in August 2020 (around the same time as the OSM import code). The feature detection code will be owned by the university and licensed to our company. The exported feature data and code to import detected features into OSM will be open source.

So it would be used to import data of unknown, possibly low, quality. With dataset unpublished and now only developed. And tightly coupled to a proprietary data format from program unusable by OSM community Mateusz Konieczny (talk) 17:49, 18 May 2020 (UTC)

The data that will be imported into OSM will obey the import guidelines. Due to the use of multispectral imagery on the drone, the detection is expected to be near-accurate (based on test images) and thus the OSM data will be of high quality (accurate attributes and correct footprints) and not low and unknown. Only the feature detection software is proprietary. The OSM import code and the data imported into OSM will be available to be used by the OSM community.

Thoughts

• OSM is not to be confused with HOT. We are close relatives, but the roles of the organisations are separate and we maintain role focus. Humanitarian projects are less interesting to me (as OSMF member) than projects that support the development of the size and skills of the OSM organisation or the more technical OSM systems that provide the platform on which others work.
• This is a interesting and cutting-edge data driven project, but it is not a community based project, or an OSM tech project. It does not reach a lot of community members. It does not build our organisation or our systems.
• I have DEEP concerns about anyone uploading AI image interpretation. There are some terrible examples in Cameroon where hundreds (thousands?) of hours of corrective work need to be done to eliminate the damage that wholesale import of AI mapping has done to the map. There is better AI work in India where urban building outlines have been fairly well traced with a few glitches. I am risk-averse to agreeing to support an AI of unknown capability that will potentially blanket whole regions with bad work. Humans work slower which gives you a chance to catch the bad mappers, contact them and teach them better skills. Full disclosure: I am professionally trained in Aerial Photo Interpretation, so I'm biased towards quality work.
  • A bit offtopic, but can you link this "There are some terrible examples in Cameroon where hundreds (thousands?) of hours of corrective work need to be done to eliminate the damage that wholesale import of AI mapping has done to the map."? Revert of an undiscussed import should be far simpler than that, and I potentially can do this (I reverted some imports violating imports guidelines already, or partially deleted some imported data) Mateusz Konieczny (talk) 12:41, 17 May 2020 (UTC)

• The success of the project is in some ways directly related to the improving the lives of people in Afghanistan. Better OSM data -> better analysis -> better results -> clear solutions for investments by donor organizations -> positive life impact.
• The imports will follow the OSM Import Guidelines and no ‘dirty’ data will be populated in OSM.
• Our organization specializes in quality control of development projects, hence the name “Development Monitors” and as part of our work we plan to implement a final-stage human-led quality control process for automatically detected features. This is also part of the proposed OSM project, which we should have included in our write up.