fix numerical stability issues in half-precision

[Devparihar5]

When using half = False (full-precision) for object detection on a GPU, you typically won't encounter numerical stability issues. This setting ensures that computations are performed with higher numerical precision, reducing the risk of instability during both training and inference. As a result, object detection on recorded videos is less likely to suffer from detection failures or inaccuracies caused by reduced precision. However, it's important to note that full-precision calculations may be slower and consume more GPU memory compared to half-precision, so the choice between the two options depends on your specific requirements and hardware capabilities.