Let's say there is a dataset consisting of photos of cars from various brands, and we're trying to train a ConvNet to identify the brand from a photo, just like these ones:
One approach I was thinking of is a two-stages approach:
- isolate, for each brand, some distinctive tell-tale landmarks (e.g. the double radiator grill for BMWs or the headlights for Porsches)
- pre-train the network with these photos only at first
- once the network gets good accuracy for the above, then train it on photos showing entire cars
The set of first-stage-training photos would look like the following:
In both stages, the output variables would be exactly the same (the car brands). The difference is only that we train with one kind of photos at first, and then use a different kind of photos.
Remember that the ultimate goal is that the network is able to tell a car brand from photos showing the entire car. We don't care about its ability to identify a particular headlight from another, it's just an intermediate step along the process.
- is this approach common for image classification?
- is there any literature on the topic?
- is it any useful or is it better to let the network learn by itself?
- should we simply mix some feature-centered images together with entire car images in one single training stage instead of separating two stages, with similar results?