Dear Research Community,
I read in Dr. Joseph's 2nd paper on YOLO, " YOLO9000:Better, Faster, Stronger " that, the anchor boxes, which are crucial for the non-maximum supression stage of the algorithm are found prior to training. To this effect, they used the k-means clustering algorithm to find the most common box shapes.
I was wondering if there are other alternatives to k-means or is k-means the best even in 2021?
With best,
Azmyin
Hi.
Of course, there are plenty of clustering algorithms nowadays: k-means, Mean-Shift, Affinity Propagation, Self-organizing map, etc. You can read about the most common approaches here: https://www.freecodecamp.org/news/8-clustering-algorithms-in-machine-learning-that-all-data-scientists-should-know/. The choice of the algorithm depends on your task and a dataset. You cannot tell for sure that this method is always better than others. Some approaches can work pretty well for small datasets, others can fit model requirements on a large data array. Also, some methods are widely used due to the ability to interpret the received result according to the task or input data.
Have a look at https://www.jmlr.org/papers/volume6/banerjee05a/banerjee05a.pdf
https://ieeexplore.ieee.org/abstract/document/1674719
Artem Kramov Ahmed Rafid Hashim and Dr. Muhammad Ali thank you very much for sharing the names of the algorithm and pointing me to the reference papers. I am in the process of writing a python tool that takes YOLO formated txt files and finds anchors from the dataset. I saw in many tutorials and even in a full-fledged coursera courset that the issue of anchor box determination prior to actually doing anything else was often overlooked.
- Badges
- Science topic
- Similar topics
- Force