Use the Tensorflow Object Detection API to detect small objects in images

Question

I would like to use the Tensorflow Object Detection API to identify objects in a series of webcam images. The fast RCNN models that have been preliminarily trained in the COCO dataset seem to be suitable, because they contain all the categories of objects that I need.

However, I would like to improve the performance of the model when defining fairly small objects in each image. If I understand correctly, I need to edit the anchor scales parameter in the configuration file so that the model uses smaller bounding fields.

My questions:

• Do I need to re-model the model in the entire COCO dataset after I configure this setting? Or is there a way to change the model for output only and avoid any retraining?
• Are there any other tips / tricks for the successful identification of small objects, without cropping the image into sections and performing output for each of them separately?

Background Information

I am currently feeding 1280x720 model images. At about 200x150 pixels, it's hard for me to detect objects.

Answer 1

• You will need to relearn completely, unfortunately, since the scales depend on the shape of the anchor.

• Having a higher resolution function map should help (but slow down the process), therefore changing the function extractor to get a smaller size (the maximum number of pools with a step> 1 usually reduces the space size) or slightly scale the image in the original image.