We collected images, labeled the data and trained our model all on our own. We now have about 330k images (about 1800 models). We used resnet152 to train it.
We are just CS students and started this just for learning and experiencing with hands on computer vision and AI. So things are not very mature. Like we haven’t make the inferencing on the end user devices. The recognition runs on the server instead. And we don’t have many images as you do. We just crawled images on the web. Also, it doesn’t work very well with cars that can’t be found at where we lived, because we dont even know those models exist lol
We had/have this problem too. If nn "sees" unknown model it tries to classify it as one of known models. To avoid this, we added 'Other/Other' bucket and put ~3k images of different cars/buses/trucks in there. Now unknown models in most cases are classified as 'Other/Other' label.