Is small size(320 or 384) of input image and one step bbox regression harmful for localization?

[xmyqsh]

Is small size(320 or 384) of input image and one step bbox regression harmful for localization?
If so, how to overcome it?

I notice that your fc6_conv do a big favor on the receptive field or content info, which will merge into fine featuremap.
And image pyramid instead of feature pyramid(roi-pooling).
And Objectness Prior to overcome the imbalance of object samples.

[taokong]

@xmyqsh
Input image with small size decreases the detection accuracy, but it makes the detector much faster. As for location, I believe one step bbox regression could also get competitive results. If you are more concern about accuracy, you could make the input size larger, such as 448/512.