If you converted the model to the SavedModel format and you noticed your model predicts fine on the CPU, but only predicts the first image/frame on the GPU, then when converting:
Run the script on a machine with a GPU (os.environ["CUDA_VISIBLE_DEVICES"] = '0')
Use the docker image tensorflow/tensorflow:2.3.0rc0-gpu (you need to use tensorflow 2.3.0rc0-gpu)
After that the prediction worked both on CPU and GPU and for all images/frames (not only the first one).
If you converted the model to the SavedModel format and you noticed your model predicts fine on the CPU, but only predicts the first image/frame on the GPU, then when converting:
After that the prediction worked both on CPU and GPU and for all images/frames (not only the first one).