I have a general question regarding retraining the SSD_mobilenet_v2.
I have downloaded 30k images of the following classes: car, truck and bus from open images, which i am using to retrain the model on jetson nano (i know it is too much…). After 2 epochs i have halted the training to check the performance. the majority of test pictures were correctly labeled and boxed.
But a good amount of them have double boxes as per below.
Why do I get this behaviour? is it due to too little epochs, too little images? possible mislabeled training data?
How would i get read of this? in post processing? do i retrain the network?