Thanks for your comment. I think object_detection_task and object_center_params are the params for centernet config, so I used them for centernet but not the others. Re the input size, it could happen to slower the learning speed, depending on the other conditions such as GPU, memory and parameters as well though. I don't see much gain in the inference performance when increasing the input size and think it's better to use more resized input images as as long as the smallest object you try to find can be represented in the resolution or unless you have no limitation in both training and inference time/hardware.