To me it seems like the right amount of data to send almost everything the driver is doing in 6500 miles over the 4-5 months. Unless we are all thinking they are not smart enough to compress video and raw data.
Again you can't train the system by collecting raw video data and simply feeding it.
Mobileye for example has 600 labeling data for them.
Tesla uses the modular DNN approach like everyone else. So for example they have a deep neural network that detects cars and a separate DNN that detects pedestrians. How deep learning works is that you have millions of pictures of cars that have been human labeled. Each picture has been cut and trimmed to only have that car so images can for example be 124 by 124 pixels. You feed those pictures to a model running on a datacenter server and that model now knows how to detect a car if you showed it.
You can also have models with different category detection. For example pedestrians, car, van, truck, traffic sign, traffic sign.
For training, Each category would be a different folder and each folder would have about one million pictures of that category. Each folder name will then act as the label for each picture when they go into the model. You then train your model using a database server cluster. Now your model learned how to detect all the above categories.
However, using individual models for each category would lead to better accuracy as your layers would only have to learn from what it needs.
For developing a tight bounding box, you need more input variables than just picture and label. You draw a box (can be 3d) representation the hieght width and length of car depending on it's orientation. This is then translated to numerical values for each picture and feed into a model.
This is how you come out with 3d bounding box around cars.
Same with distance of objects and recognition and classifying objects, image segmentation. Everything MUST be labeled.
They are clearly using the onboard supercomputer processors in the car to make decisions about what to do in certain circumstances so it would also make sense that they could use it to decide what things it does not understand and mark that raw data to be saved long enough to transmit it over Wifi back to the mother ship, though based on the 1-1.5g per day for a guy who drives an hour a day seems about right for compressed video/image and data from 8 cameras and sensors.
A computer doesn't understand what it doesn't understand as it pertains to Vision. All it sees are pixels which are numbers RGB 0-255.
So it can't look at an image and say well that's a car i haven't seen before because it won't know what it's looking at is a car or anything all.
