I have trained a football detection model. I have so far trained the models using RCNN, SSD (backbone MobileNet), CenterNet and others. SSD and Centernet, so far have been the best in terms of speed and accuracy. I am using TfLite versions of those models for my Camera based android application. My use case involves the detections from the models and some business logic using those detections. The problem that I have been facing is that, quite often other objects get detected as footballs. For example, as the only feature in a football is its roundness, objects like Clock, Round Lamp, Fan etc get falsely detected as footballs. And sometimes, my white shoes tend to get detected as football a lot of times too. This horribly messes up my business logic and therefore, the user experience.
I have managed to get a dataset (with augmentation) of about 48k images. The models are pre-trained on COCO dataset.
Now, I am not so sure if I will ever be able to get 100% accurate football detection model, or even close to that. But I was wondering there might be some post-processing stuff that could help me in mitigating mess-ups that are caused by false-positives. For example, kalman filters, or inferring the next location of the ball using some simple linear interpolations using previous 15 predicted ball locations.
Could anyone here share some ideas in sorting out my problem ? Would really appreciate it.