Finetuning a tensorflow object detection pretrained model

Question

I'm working on a real-time object detector with tensorflow and opencv.

I've used different SSD and Faster-RCNN based frozen inference graphs and they almost never fail.

The video stream comes from a IR camera fixed to a wall that has a background that almost never changes. There are some misdetections at particular hours of the day (eg when the light changes in the afternoon) that occur in the background area or on small objects too close to the camera.

So to fix this little mistakes i wanted to finetune the model with images from the same background.

Being the background always the same, how do i approach the retraining of the model having 1000 misdetections pics that are all almost the same?

Answer 1

In case of variations in the background lighting, it might be possible to use Background Subtraction

https://docs.opencv.org/3.4.1/d1/dc5/tutorial_background_subtraction.html ,while dynmically updating it as shown here:

https://www.pyimagesearch.com/2015/06/01/home-surveillance-and-motion-detection-with-the-raspberry-pi-python-and-opencv/

Thank you.