Hello,
I am searching for literature on methods employed for choosing object classification when multiple object classifications (e.g. labels) are given with relatively high-confidence during analysis of a live video feed frame by frame (e.g. frame0:truck, frame1:sedan, frame2:sedan, frame3:truck)?
For instance, say a large truck enters the field-of-view of the video camera and is detected. The image segment within the provided bounding box from the detection algorithm is classified as a sedan with 95% confidence in the first frame, then subsequently classified as a truck with 95% confidence in the following frame. This process continues, resulting in multiple object classifications (i.e. labels) that vary between sedan and truck with high confidence values. How can I select the correct optimal classification? Are there methods that utilize all the classification (labels) of the object to select the optimal classification?
Hi Ian Robert Stewart,
I Would suggest you use a non-maximum suppression technique iterating over bounding boxes from the frame over a short time period. consider the first 5 frames from your example when you run non-maximum suppression create a custom sorting metric to use the number of occurrence and score to pick out the best class and bounding box and keep moving to the next 5 frames. (reference for non-maximum suppression https://www.pyimagesearch.com/2015/02/16/faster-non-maximum-suppression-python/). If you are facing the same issue with moving objects it's gonna be difficult to do this.
Hello,
I wrote about this a while back and am still studying about it, hence why my preprint is private. However you could take a look at this link and see if you get a clearer idea of it. Experiment Findings Skip Network In a Video
- Badges
- Science topic