Use Tensorflow Object Detection API to detect small objects in images

Question

I'd like to use the Tensorflow Object Detection API to identify objects in a series of webcam images. The Faster RCNN models pre-trained on the COCO dataset appear to be suitable, as they contain all the object categories I need.

However, I'd like to improve the performance of the model at identifying fairly small objects within each image. If I understand correctly, I need to edit the anchor scales parameter in the config file to get the model to use smaller bounding boxes.

My questions are:

• Do I need to re-train the model on the entire COCO dataset after I adjust this parameter? Or is there a way to change the model just for inference and avoid any re-training?
• Are there any other tips/tricks to successfully identifying small objects, short of cropping the image into sections and running inference on each one separately?

Background info

I'm currently feeding 1280x720 images to the model. At around 200x150 pixels I'm finding it harder to detect objects.

Giavani

Answer 1

1. You'll need to retrain completely unfortunately, since the weights do depend on the shape of the anchor.

2. Having a feature map with higher resolution should help (but slow down the process), so changing the feature extractor to get one with less input size reduction (max poolings with stride >1 is usually what reduces the space size) or upscaling the image a bit in the initial image resizer.

Thimbleweed

Thanks - since I am going to need to re-train is there any benefit to using a training set that includes plenty of examples with small objects? Or is the classification part of Faster RCNN sufficiently independent from the bounding box detector to allow good performance to be achieved regardless, assuming I adjust the anchor settings? – Giavani

I'm not sure I understand, but you'll have to use images with small objects to train the network to detect small objects (adjusting the anchor settings will make it easier to detect small objects, but you'll still need to ave some in your training data otherwise it'll just learn that there are only rarely objects corresponding to the small anchor sizes) – Thimbleweed