We're releasing two pre-trained resnet50 classification models here.
We trained them on an internal dataset we were collecting
- between the months 2019-05 and 2019-11
- mainly 8am -- 6pm with some outliers
- mostly daylight, in later months also dark / low light conditions
We used the nearest neighbor index to cleverly bootstrap and annotate everything construction / blockages related automatically which either directly influences the traffic / vehicle behaviors or influences other people's road behavior (think: there's a construction site on a bike path, so bikes have to take the road for a bit). We then manually validated the results and iterated multiple times on our internal dataset.
At this time we can not release our internal dataset, but we hope this construction / blockages model can be of help on its own for similar use cases and scenarios.