Thanks a lot for the discussions and your time. I will definitely try some quality control methods as you suggested. Sometimes more data also increase the accuracy even when it's noisy, that part is not an exact science unfortunately:) But I will try using only %100 validated tasks as you suggested. We might also try to use the algorithm we are developing for quality control, inspect the tiles that the algorithm and the current labels disagree too much, discard some of them where the labels seem to be inaccurate and train the AI model again. We can see and compare the results and accuracy afterwards. It seems that the quality will be a real concern, thanks for raising the issue and your suggestions.
On Tue, Nov 20, 2018 at 9:55 AM Pierre Béland <pierz...@yahoo.fr> wrote: > As John is saying, quality can vary a lot. You should look at the thread > where I presented geometry quality analysis recently. > https://lists.openstreetmap.org/pipermail/talk/2018-September/081392.html > <https://urldefense.proofpoint.com/v2/url?u=https-3A__lists.openstreetmap.org_pipermail_talk_2018-2DSeptember_081392.html&d=DwMFaQ&c=slrrB7dE8n7gBJbeO0g-IQ&r=9jC6quPZN0RnI866qaQhEg&m=CWFFLleRsuBiOUXD6gXRpreKxzA-4fxMrEBXvJSZsDo&s=Fas5ZTiyu-liuodzbu_RkBxMU0iiNHPdhZyqE85GT4Q&e=> > > You need a more rigourus procedure where you can compare AI with > contributors that operate in the same conditions, with the same imagery. > > > Pierre > > > Le mardi 20 novembre 2018 09 h 44 min 59 s HNE, John Whelan < > jwhelan0...@gmail.com> a écrit : > > > I've done a lot of validation in HOT. Some is done by people with little > experience so I don't think validated tiles would be much use to you. Also > be aware that some imagery can be three years out of date so the imagery > can vary. The sort of problems you'll run into are much of the HOT mapping > is done by inexperienced mappers using iD and the buildings are more > approximate than you might like especially on size and shape. It takes > four times longer to correct a building than to remap it so as a rule of > thumb validators tend to stay away from projects with buildings. > > Having said that have a look at the hot projects for one that is 100% > complete and 100% validated. You should be able to work out the > coordinates from the project and if you are lucky it should also give you > the imagery used when mapping for HOT. Beware this OpenStreetMap and some > mapping may have taken place on the ground showing buildings etc that are > not on the image. > > You might be better off working with an experienced mapper who can both > map and check the sources. They should also check for duplicates, there > are tools available. I suggest an area that hasn't been mapped before and > map it correctly with an experienced mapper using JOSM and things like the > building_tool plugin. > > If you're just trying the algorithm out Ottawa in Canada has accurate > buildings in OpenStreetMap. > > Have fun > > Cheerio John > > Georgy Potapov wrote on 2018-11-20 9:25 AM: > > Dear John, > > the question isn't about imports. Seems guys are preparing dataset to > train and test models and they need to download data that's been already > validated within Missing Maps, as a "ground truth". > Is there a way to get the precise coordinates of the areas where this job > is done or projected to be done? > Me too was asking if it makes sense to apply algorithm to detect damaged > buildings on imagery to leverage the mapping speed in disaster affected > areas? > > All the best, > Georgy > > On Tue, Nov 20, 2018 at 3:29 PM john whelan <jwhelan0...@gmail.com> wrote: > > Essentially you are proposing an import. There are OpenStreetMap rules > about how this should be done. It can be done, Microsoft has released > building outlines for the US which were created in this manner but the > import itself is being done in sections by conventional mappers. > > You will need the cooperation of the local mappers on the ground and > imports of dubious quality tend to get a fair chunk of flack. I would > suggest you Google OSM imports before doing anything. > > Cheerio John > > On Mon, 19 Nov 2018, 10:47 pm Serkan Karakulak <sk7...@nyu.edu wrote: > > Hi, > I am a graduate student in a Data Science program, and I wanted to get in > touch with you because me and my two other friends are interested in > working on a machine learning project to map areas using their satellite > images and produce their labels. If we obtain a high accuracy, we thought > it could be of use to the hotosm and the missingmaps initiatives. > > I have first heard about Missing Maps when I came across a mapathlon event > two years ago in Istanbul. We were in search for a class project for one of > our courses and I remembered about the mapathlon event and we would be very > happy if we could contribute to the this wonderful initiative. There are > already some previous successful works on this subject, so we are very > hopeful that we could come up with an algorithm with high accuracy. > > In order to develop the model, we need to extract bounding boxes of the > validated tiles we see at hotosm. Then we will extract the satellite images > and their labels using the label-maker API which is developed by DevSeed. > Is there a method to extract the coordinates of these validated tiles? > > Thank you for your time and help. > > Best, > > Serkan > _______________________________________________ > HOT mailing list > HOT@openstreetmap.org > https://lists.openstreetmap.org/listinfo/hot > <https://urldefense.proofpoint.com/v2/url?u=https-3A__lists.openstreetmap.org_listinfo_hot&d=DwMFaQ&c=slrrB7dE8n7gBJbeO0g-IQ&r=9jC6quPZN0RnI866qaQhEg&m=CWFFLleRsuBiOUXD6gXRpreKxzA-4fxMrEBXvJSZsDo&s=7yKuZLIGwWnrZsIY7HCqPZ_3BagzqQFFGCmM5TcVFoc&e=> > > _______________________________________________ > HOT mailing list > HOT@openstreetmap.org > https://lists.openstreetmap.org/listinfo/hot > <https://urldefense.proofpoint.com/v2/url?u=https-3A__lists.openstreetmap.org_listinfo_hot&d=DwMFaQ&c=slrrB7dE8n7gBJbeO0g-IQ&r=9jC6quPZN0RnI866qaQhEg&m=CWFFLleRsuBiOUXD6gXRpreKxzA-4fxMrEBXvJSZsDo&s=7yKuZLIGwWnrZsIY7HCqPZ_3BagzqQFFGCmM5TcVFoc&e=> > > > -- > Sent from Postbox > <https://urldefense.proofpoint.com/v2/url?u=https-3A__www.postbox-2Dinc.com_-3Futm-5Fsource-3Demail-26utm-5Fmedium-3Dsiglink-26utm-5Fcampaign-3Dreach&d=DwMFaQ&c=slrrB7dE8n7gBJbeO0g-IQ&r=9jC6quPZN0RnI866qaQhEg&m=CWFFLleRsuBiOUXD6gXRpreKxzA-4fxMrEBXvJSZsDo&s=QhPk-pKNdhLgkaPqpfJM_wmPGO5FfXIiEEj0eoZ2QME&e=> > _______________________________________________ > HOT mailing list > HOT@openstreetmap.org > https://lists.openstreetmap.org/listinfo/hot > <https://urldefense.proofpoint.com/v2/url?u=https-3A__lists.openstreetmap.org_listinfo_hot&d=DwMFaQ&c=slrrB7dE8n7gBJbeO0g-IQ&r=9jC6quPZN0RnI866qaQhEg&m=CWFFLleRsuBiOUXD6gXRpreKxzA-4fxMrEBXvJSZsDo&s=7yKuZLIGwWnrZsIY7HCqPZ_3BagzqQFFGCmM5TcVFoc&e=> > _______________________________________________ > HOT mailing list > HOT@openstreetmap.org > > https://urldefense.proofpoint.com/v2/url?u=https-3A__lists.openstreetmap.org_listinfo_hot&d=DwIGaQ&c=slrrB7dE8n7gBJbeO0g-IQ&r=9jC6quPZN0RnI866qaQhEg&m=CWFFLleRsuBiOUXD6gXRpreKxzA-4fxMrEBXvJSZsDo&s=7yKuZLIGwWnrZsIY7HCqPZ_3BagzqQFFGCmM5TcVFoc&e= >
_______________________________________________ HOT mailing list HOT@openstreetmap.org https://lists.openstreetmap.org/listinfo/hot
