Detection results are worse for smaller images than those specified in pipeline.config

Hi i have a question,

I use an SSD Mobilenet V2 network for detection. I have found that the result is worse if my input image has a smaller image size than what is specified in the pipeline.config. For bigger pictures, that would make sense to me. Is there a logical explanation why I get a poorer quality with smaller pictures? A resolution of 768x320 is specified in the pipeline.config and I tested it with images with 738x119.

Have you verified that the performance worse for small objects?


I think I explained it too imprecisely. The object to be recognized is exactly the same size in both images. Only everything “uninteresting” is cut away. That’s what amazed me.

Do you have an image result to post here with the object of the same size but at different resolution?