Hi @Morganh
Your suggestion of changing the target class mapping to vehicle → vehicle and all other instances of car to vehicle worked, but I don’t understand why. TrafficCamNet expects a label of car according to the labels it was trained on, so as far as I can tell this label translation is not provided in any of the spec files. How does TCN know that a vehicle is a car without providing a target class mapping indicating this?
One more thing. After implementing the changes you suggested, I noticed that TCN was producing poor inference results similar to those described in this ticket. So I added load_graph: true in the model spec and that considerably increased validation performance.
Thanks for your help.